FIELD GUIDE · TARKOVSKY DIRECTOR
Tarkovsky Director:
see what it decided for you.
Tarkovsky Director directs and produces the film within the budget you set. The kickoff asks, once and before any paid step, what it may decide for you. It then works through the same routes, checks and prices as the page, never loves an unchecked still, never waves a must-fix flag through, and reports every decision with where it came from and what it cost.
Watch this stage in the video tutorial →
12a / TARKOVSKY DIRECTOR
Leave it to Tarkovsky Director, and see what it decided for you.
Tarkovsky Director in 0.68.5: a still remade on another model
- A still whose pictures fill its model goes without the sketch, and if its still check then finds the framing wrong, Tarkovsky Director remakes it once, within its remakes, on an image model that takes all the pictures and the sketch. Its report lists that frame as remade on another model.
- That still is priced, held and charged on the model it was made on, as if you had picked that model for it, and so is an angle of a person or a thing made first for it. A model that takes fewer pictures than the frame has is never used for it. Every other still goes on the project's image model, as before.
Tarkovsky Director in 0.68.5: calls at once, helpers on long films, and lighter turns
- A turn's calls at once. When one turn asks for several things that don't depend on each other, the studio does them together: reads at once, and a change waits for the calls before it. The answers go back in the order Tarkovsky Director asked, so it decides on the same answers as before, sooner.
- Helpers on long films. On a film of four scenes or more, Tarkovsky Director may hand some scenes to a helper: the same model, run and budget, working at the same time. A helper changes only its own scenes and reports back when its task is done. Up to two work at once. While a helper works, Tarkovsky Director leaves its scenes to it: it doesn't change them, and stills or takes it asks for the whole film skip them. The panel shows each helper's step and scenes, and each of its lines in the log starts with its name. Stop stops the helpers too, and Resume carries them on.
- Lighter turns on long runs. When a long run's conversation is full, it now lets go of its oldest turns several at a time instead of one every turn. So most of each turn is the same as the turn before, and the model reads it from its cache: faster and cheaper. It always keeps its last four turns whole.
- After a restart, each turn that was in flight waits for review in Admin, as one turn did before: with two helpers working, that can be three turns, Tarkovsky Director's and each helper's. Each is taken again, so a turn settled later as charged is paid twice.
Tarkovsky Director in 0.68.3: a run pauses when the studio can't reach its AI models
- When the studio can't reach its AI models, every page says so in one line: The studio can't reach its AI models right now. This is on our side, not yours. Nothing was charged. Please try again later. It isn't something you did, and the call costs nothing.
- A Tarkovsky Director run pauses instead of failing. Its panel says Paused, and nothing is charged for the turn it couldn't take. When the studio is back, open Tarkovsky Director and press Resume. The run carries on from where it stopped.
- Everywhere else, such as a chat, a still or a take on the Storyboard and the Shot List, or Story's script suggestions, the same line, or a short form of it, says what happened.
Tarkovsky Director in 0.68.1: stills scene by scene, and coverage drafts side by side
- Stills scene by scene. When a run goes as far as stills, it makes scene 1's master still and its other stills, each with its still check, before it covers the rest of the film; then each scene's stills follow its own Check shot list. So a long run shows pictures early. It makes the same stills as before, in another order.
- Coverage drafts side by side. When one turn drafts coverage for several scenes, the studio drafts up to four at once (two drafts of one scene one after the other). Each is quoted and held from the run's budget as before, and together they never pass it.
The agent path in 0.68: Key frame 2 and the frames between, and a camera placed again
- A frame by its number.
kfon the camera, start, pose and move routes takes a frame's number (2 for Key frame 2). Through/opsa frame by its number has its own ops:frameCamera,frameLens,frameStart,frameMoveandframeFigure. - A camera placed again.
{ kind: "place", shotId }on/opsplaces a shot's camera again from its frames, by the Shot List's own rule, and answers which frames it placed and which it left (200 with nothing placed when there is nothing to place).
Tarkovsky Director in 0.67: the Dailies Critic's new prompt, and a render at your size
- The Dailies Critic watches with a new prompt since 0.67 (Dailies has what changed). So when a run starts or resumes, the Critic first looks again at each take in the cut it watched with the earlier prompt, as for any take watched with an earlier version: each watch priced first and paid from the run's own share of its budget, about 2 to 4 credits a take, so about 40 to 110 credits on a film of 20 to 30 takes. A take a run doesn't reach keeps its earlier notes until it is watched again.
- A render asked without a size is made at the size saved on Export; a run that names a size gets that size.
- Where the story stands: a state a run writes on a shot reads in the studio's line where its words allow, with Accept and Edit for you. What carries on from shot to shot is on the shot panel; a run reads and writes each shot's own state, as before.
The agent path in 0.67: cut offers, story state, a render's size, who the Critic keeps out
- Cut offers.
GET /api/projects/:id/plan-cuts/offers(free) lists the cuts the shots' words ask for that no mark answers, each with its shot, part, second and kind.POST /api/projects/:id/plan-cuts/acceptwith{ revision, shotId, part, at, kind }marks one (free; a History line with Undo; 409 on a stale revision, 400 outside the part or past six marks). - Story state.
shots[].storyStateis still each shot's own rows: carrying on is read, never stored, and the studio's proposals are on the page, not in your read of the project. A row withadds: trueadds its words to what carries on from the shots before (with no words, it keeps what carries on); a row with no words and noaddsends the state there. Set a row where the story changes someone. - A render's size. A
renderjob withoutfinishing.resolutionis made at the project's saved size (720p or 1080p); one with it, at that size. - The Dailies Critic. A watch's
keptOutlists only who Frame it's floor plan has out of the scene at the take, and is empty without a floor plan. Watches made since 0.67 carry the new prompt's version inpromptVersion. - The narrator.
POST /voice-oversends an ElevenLabs voice a 0.4 s pause after each verse line a Final Draft, Fountain or Word script sets on a line of its own.GET /voice-overgives a line's words as before, a\nat each verse break. - A product's pictures. Its white packshot line and its side or back line are put back after your words when you leave them out, as a thing's are.
Tarkovsky Director in 0.66: the Dailies Critic looks again when a run starts or resumes
- When Tarkovsky Director starts or resumes, the Dailies Critic first looks again at each take in your cut that it hasn't watched, or watched with an earlier version of its prompt, in the film's order. Each watch is priced first and comes out of the run's own share of its budget, never the credits you kept for takes or for fixes; the estimate counts them before you start (Dailies Critic, looking again).
- The run waits for them, 15 minutes at most, then carries on with what the Critic says now. A take it couldn't watch keeps its earlier notes, and the run's log says why.
- A note you answered on an earlier watch isn't carried over: if the new watch raises it again, it asks again.
- Looking again has limits. It keeps enough of the budget for the run's own next turns, so on a small budget it may look at few takes or none, and says so with the figures. It looks only at takes already in the cut when the run starts or resumes: a take cut in later waits for the next Resume. After 15 minutes the run carries on, and it can finish while takes it hasn't reached still read by their earlier watch; its log says how many. You can ask for any take's watch yourself in Dailies.
- Before it finishes, Tarkovsky Director waits for the Dailies Critic's watch of any take in the cut that is still being watched.
Tarkovsky Director in 0.66: Gemini 4 Carbon, the Critic's model, and what it is told
- Gemini 4 Carbon is among the models that can run Tarkovsky Director. It takes video, so it can also be the Critic.
- The Critic is Gemini 4 Argon unless another is chosen. A run that Gemini 4 Argon itself directs is judged by the next model that takes video, as before.
- It writes the camera a stretch at a time: each stretch a hold, or one move with where it ends.
- It lays one ambience bed for each scene, one place at one time. When the place or its time changes, it makes a new bed.
- Take the best version. Only when your kickoff turns it on, and a still's or a take's remakes are used up, it may go on with the best version. It gives its reason, and its report lists it as taken at its best, never as fixed.
- When a model keeps getting a prop wrong, it describes the prop as it should read on screen and frames it so the wrong reading can't show.
The agent path in 0.66: the floor plan's question, a group, Reply, the shots beside each other
- A take of a shot that doesn't match the floor plan yet is refused once (409
plan_follow), with the question and how to answer it:POST /api/projects/:id/viewfinder/:sceneId/opswithops: [{kind: "follow", shotId, answer}],answerone oftake,matchorkeep, then ask again. A Tarkovsky Director run is told the same. - A still check's
issues[].fixreads in the new words from 0.66 on; a note filed before keeps its words. A group's count note is raised, and stands, only for an exact number. - A group's sheet is one picture. Make its front with
asset_image; a still or a take that names the group is sent it, with one sentence that says about how many the group is. The switch and the number arePOST /api/projects/:id/assets/:assetId/group({ group, groupCount }, 2–500, free). The other angles of a group are refused (409). - Reply on a move's end:
POST /api/projects/:id/storyboard/move-end/reply/quotewith{shotId, ask}is free;POST …/move-end/replywith{shotId, ask, maxCredits, model}is paid and answers{reply, propose, ends, chargedCredits}. Accept a proposal with the Viewfinder'smoveEndop andon: {words}. - Go to Edit's cut is made without Loved takes only for a Tarkovsky Director run and an agent: the tick is the filmmaker's own.
- The shots beside each other.
GET /api/projects/:id/pretake/continuity-criticlists each pair the film has;POST …/quote { takeId }prices comparing that take with the settled take of the shot before it, andPOST … { takeId, maxCredits }runs it. Each answers 409 while the comparison is switched off in Admin. - Publish to YouTube is unchanged for an agent:
POST /api/gallery/youtube/uploadstill needsconfirm: trueand stays the filmmaker's own decision.
Tarkovsky Director in 0.65: a retake keeps what works, and names what a fix replaces
- When Tarkovsky Director makes a shot again after the Critic's notes, the take's words carry each fix and then what worked ("What to keep: …"). Where it names the plan's words a fix contradicts, those words are replaced for that take only, exactly as named, as your note on a take does, and the run's log says so ("… replaces … → …, this take only"). Otherwise each fix goes beside the plan's words, and the log says it was added (or left out, when the take's words were full).
- Before it finishes, the export must be of the cut as it plays now.
- Change a clip the Critic trimmed and that take keeps your choice: no cut built later trims it to that stretch again.
Tarkovsky Director in 0.65: the narrator reads verse as verse
- Verse is read as verse. Where the script marks a verse break, the narrator is sent a line break, not the slash. The script's own " / " always counts. A poem's lines on lines of their own count in a script from Final Draft, Fountain or Word (two lines or more, each starting with a capital or following punctuation), and in one typed in Story (three lines or more), never in one imported from a PDF, whose lines are where the page wrapped them. Lines made with the slash before stay made. ElevenLabs reads a line break as a space, so until 0.67 a verse line paused only where its punctuation did; since 0.67 an ElevenLabs voice pauses 0.4 s at the end of each verse line a Final Draft, Fountain or Word script sets on a line of its own (Edit has the details).
The agent path in 0.65: story states, a retake's replace, the narrator's lines
shots[].storyStateis each person's state at a shot ([{tag, state}]). What a Director run or an agent writes there is a proposal the filmmaker sees and can keep or change; the server stamps who wrote it. A row the filmmaker wrote is never changed or removed by your save: your words go beside it as a proposal. Leave the field out of a save to keep what's there.POST /takes/:takeId/remake(and its quote) takesoverrides: [{ kind, ref, to }], a take note's overrides, up to 20, sent to every remade part. Tarkovsky Director'smaketakesreplace: [{ was, now }](RT17).- GET
/voice-overgives each line as sent: a\nat each verse break.
Tarkovsky Director in 0.64.1: the Critic on Dailies watches its takes, and it answers to it
- Every take it makes is watched as it lands (Dailies has the details). Each watch is priced first, counts in the run's studio spend, and stays inside the run's budget less the credits kept for fixes. A watch that would pass the budget is skipped. The estimate before a run has a Dailies Critic line.
- It hears each watch on its next turn, with every note's time, weight and fix, and what works. A watch that didn't run says why. Older must notes on the takes it loved are told once, at the start.
- It waits for the watch. Its love or like of its own take is refused until that take's watch lands (at most 15 minutes after the take did), and a must note is a must-fix flag it can't love over.
- Its retake carries the fix. When it makes the shot again, each open must note's fix goes after its own words as the new take's direction.
- It doesn't deliver over a must. Finish is refused while a take in the cut has an open must note, unless you accepted it, set it aside as Not a problem, or Take the best version took it (the report lists those). When it can't fix one (its remakes used up), it flags it for you and the run waits for your answer: accept it, or tell it to fix it.
- The kickoff decides it per film. Answer No to Should a separate model watch each take as it lands, and hold it to its plan? in the Critic's question, and no take on that film is watched by itself. The estimate has no Dailies Critic line, and the run's first message says The Dailies Critic: No. It's your answer: Tarkovsky Director and outside agents can't change it. No answer reads as Yes.
- The Admin page: the switch at
/api/admin/dailies-criticis on by default.POST { "on": false }turns it off for everyone, and every flip is in the audit trail. The server'sDAILIES_CRITICsetting (onoroff) decides over it.
Tarkovsky Director in 0.64: five checks on what it waves through
- A take is judged after it's seen. A look at a take reaches Tarkovsky Director with its next turn, so a love or like of a take it asked to see in the same turn is refused until it has looked.
- A take's plan covers its length. When a shot is lengthened, for its narration say, Tarkovsky Director's take waits until the shot's timeline or camera lines say what happens in the new seconds, or that the frame holds. Your own takes in Dailies are never held for this.
- No still warnings on a film without stills. When the kickoff makes no stills, Check shot list no longer raises the notes about a still's own references. Takes don't use them, and so many of them taught it to set warnings aside.
- A place's plan isn't approved over its last look. While the plan's last look lists a change the studio couldn't make (a fallen slab over a hollow, say), Tarkovsky Director's OK waits until it makes the change and says what it did, or redraws the place. Yours goes through as before.
- An off shot can't lean on another off shot. When Tarkovsky Director or an agent switches a shot off "because 1V carries it", 1V must be on. When it switches off a shot another off shot's reason leans on, its answer names the beat nothing shows now.
- The Admin page: the Critic on Dailies has its switch at
/api/admin/dailies-critic(off in 0.64, on since 0.64.1: above).
Tarkovsky Director in 0.63.11: the Critic's notes land where it saw them
A Critic pass could write its times in minutes and seconds, 3.5 for 3:50, and the studio read them as seconds, so every note of a long film fell on its first shot. A pass that writes its times that way is now read that way, and each note sits on the shot it saw.
Tarkovsky Director in 0.63.10: the Critic watches a whole film
A cut longer than about three and a half minutes was refused before the Critic could watch it: the copy it watches was too large to send, and nothing was charged. The copy is smaller now, so the Critic watches a film of up to 8 minutes. A longer film still gets the free measures.
A new run on a project that has had one before, and a Resume, now always reach Tarkovsky Director with your note. On a large project they could start before the studio had written it, and Tarkovsky Director answered that it had no project.
Tarkovsky Director in 0.63.9: sound counts against its budget
Sound effects, ambience, music and set packs a Tarkovsky Director run makes now count against its budget, as its stills and takes do. They were always charged to your credits. Now the run's total shows them, and it stops at its budget when it makes them.
When shots carry a speech in pieces, Once, in Edit makes those pieces and only the words no shot carries.
Tarkovsky Director in 0.63.8: what the kickoff can ask for
Four new kickoff answers. Each is yours alone, and each is off unless you turn it on.
- Takes with no stills. On the camera from Frame it, Tarkovsky Director makes no stills: the take follows the gray blockout. A model that needs a first frame can't go without them.
- Where a voice-over is made. Once, in Edit makes each line once, in the character's voice from Assets, and lays it on the narration track. The takes stay free of narration. It's priced first.
- Credits kept for fixes. The Critic watches the cut, and these credits pay for what it finds. A retake of a shot it named can go past your remake limits while they last. Its second look says what was fixed, and the report lists each note. If the figure is too low to pay for one round of fixes, the kickoff says so and gives the number. You can still start.
- Take the best version. When a still or take has used up its remakes and a must-fix flag still stands, Tarkovsky Director may go on with the best version. It says why, its report lists each one, and you can still answer the flag.
Fix now works when your remakes are set to Only flag it: it opens one remake of the still or take the flag is on, within the run's budget. Answer Fix again and it opens another. A take's flag stays open until a check of the new take clears it.
On the camera from Frame it, a shot of up to 15 seconds is one take with no join, so a note on any part remakes the whole shot.
A model that can't watch video sees a take as frames: first, middle and last, about one a second on a contact sheet with their times, and the words heard in it. Gemini 4 Argon watches the video itself.
When the budget is tight, a turn goes with a shorter answer. If no take can start, Tarkovsky Director has a few short turns inside the credits kept for takes to clear what they wait for.
Tarkovsky Director in 0.63.6: its turns stop growing, and credits stay for takes
- Credits kept for takes. A number on the kickoff's budget question, empty by default. In a Tarkovsky Director run, its own thinking stops before those credits; stills, takes and checks can still use them, and each credit they spend comes off what is kept. Tarkovsky Director can't change the number. A budget has to be more than the number, or the run won't start.
- When its own credits run out. If credits are still kept, the studio makes one take of each shot that has none with them, with no more thinking and nothing remade, as fast as the three lanes allow. Nobody judges those takes, so a run that ends this way leaves them for you to love. If a take can't start (the board isn't locked, a frame has no loved still, a question is open), the run stops, says why and says how much is unspent; do what it says and press Resume. Once every shot has a take nothing is kept any more, and what is left goes back to the run.
- Long runs cost less a turn. Tarkovsky Director's memory has a ceiling: past it, the oldest turns leave its conversation, and it works from its own digest, its notes and the project. Each turn's cost is what the model's provider charged, so what its cache saves comes off the turn; a turn is still held at its full price first, and the rest goes back.
- It stops later. It used to set aside a whole expensive turn for its last account; now it sets aside about one and a half of its recent turns, so more of the budget is used. It still can't pass the budget: when the last account doesn't fit, it writes it from its digest alone, or the log is the record.
Tarkovsky Director in 0.63: the Knowledge Vault (TAR-606)
- What it reads. Besides its own pinned prompts, Tarkovsky Director reads the studio's Knowledge Vault on every run: first the method (how the studio directs performances on every film, read before anything else), then recipes (how the studio makes something, such as a dialogue scene), prompts (templates the studio has tested), skills (ways of working, such as Keep and Avoid for retakes) and notes (facts about a model, such as what Seedance 2.5 refuses). When it follows a recipe, its decision log names it with its version. Its own rules (your words, money, access and the checks) come first: when an entry disagrees with one, Tarkovsky Director follows the rule and says so in the decision log.
- The version a run reads. A run reads the vault as it was when the run started, for its whole life, Resume included; the run's first log line says which version (It reads the Knowledge Vault, version 3.). A run started after a change reads the change.
- For admins: Admin › Tarkovsky Director › Knowledge Vault. Read each entry exactly as Tarkovsky Director reads it; Edit it (say in a line what changed and why: it goes in the history); History lists every version with who saved it and when, compares any two line by line, and Bring this back makes an old version the newest again. Archive takes an entry out of what Tarkovsky Director reads; Put back returns it. New entry adds one. The two recipes and the prompt that came with the release arrive archived, because they change what runs spend or which route they take: Put back each one once it has been reviewed. Nothing is ever deleted or rewritten: every save, restore and archive is a new version, and every change is in the audit trail.
- The vault's switch. Under the vault's version line: On, Tarkovsky Director reads the vault, each run the version it started on; Off, it reads only its pinned prompts, from each run's next turn, and a run started while it is off never reads the vault. Turned back on, each run reads the version it started on again. When the server's
DIRECTOR_VAULTsetting is on or off (on,true,1oryes;off,false,0orno), it decides and the switch waits; any other value is ignored, with a line in the server's log, and the switch decides. Every flip is in the audit trail, and in the log of each run it changes, at the turn it takes effect (The Knowledge Vault is switched off: from this turn, Tarkovsky Director reads only its pinned prompts.). While it is off, the version line says how many words the vault holds, not what Tarkovsky Director reads. - What Tarkovsky Director reads shows the whole text after the pinned prompts, with its word count (12,000 words at most); Vault history lists every version of the vault, with how many runs started on each. Pinned prompts shows the files Tarkovsky Director and the takes are built from (Tarkovsky Director's role, its playbook, its tools, the take and still templates), read-only: they change only with a release.
Tarkovsky Director in 0.60: every flag that waits can be answered (TAR-345)
- A flag Tarkovsky Director wrote down on its own (no note or check behind it: its check reads decision log) takes Accept and Fix like the others. Accept closes it; Fix keeps it waiting for its fix. Your answer is kept as your own line of that run's decision log, which says who answered and how (accepted it, said fix it). Until 0.60 the answer said That flag is not in this project. and the flag kept waiting in the panel and the report.
Tarkovsky Director in 0.59: what it asks you to answer is in the panel, and a link (TAR-577, TAR-578)
- Its words link to the place. In the panel's Waiting for you: line, the report's summary and What is left, each shot Tarkovsky Director names (1B, 1F) opens on the Storyboard, each page it names (Storyboard, Shot List, Dailies…) opens, the flags it asks about and the report open the run report, and the remake limits (Only flag it) open the Kickoff.
- Waiting for your answer (TAR-578). Under the panel's Waiting for you: line, Tarkovsky Director lists the must-fix flags it can't make a take over, read as they stand now, each with its frame (a link to it on the Storyboard). Every one not answered yet starts ticked: Accept lets the takes go ahead with the stills as they are; Fix sends them back for Tarkovsky Director to remake and check again. No reason is needed. Then press Resume. Fix needs remakes allowed: with the Kickoff on Only flag it it is off, and the panel says so, with a link to the Kickoff. A flag you sent back stays in the list, marked you said fix it, until its fix lands.
- The run report's Flags (Decided for you › Run reports) have the same ticks and the same Accept and Fix; a single flag's Answer is two buttons, Accept and Fix it, with an optional reason (until 0.59 it started on Fix it and needed a reason). The flags Tarkovsky Director names in its words (the 13 must-fix pre-take flags, the report) open that report at its Flags.
- A flag counts once, as it stands now. A note raised outside a run (an earlier run's, or a still check's that landed while no run was open) that a later run asked about now reads as its note: fixed once its still is replaced or checked again, answered once you accept it. Before 0.59 it stayed "asked" in that run's report and in the panel's count for good.
Tarkovsky Director in 0.57.1: it waits for you instead of ending without the film
- Waiting for you. When Tarkovsky Director can't go on without you (a red flag on a floor plan that only you can clear, or a budget that ran out), it stops and says exactly what to do. Its pill reads Waiting, and the panel Waiting for you: with the reason; the report's first section says the same. Do it, then press Resume: it carries on from where it was, with what is left of the budget or a new one. Until 0.57.1 such a run ended as Done: the panel said the film was exported when it wasn't, and the run couldn't be resumed. A run like that now reads Finished without an export, with its reason; start a new run to carry on.
- Floor plans. Tarkovsky Director approves a location's floor plan once it has fixed every red flag and every flag whose fix it can make. A yellow flag it can't fix (such as "Doesn't read as a field yet") stays on the plan for you, and no longer holds the location's stills and takes; the approval's line in History names it. A location's plates, stills and takes still wait for its approved plan, or for Skip the floor plan.
Tarkovsky Director in 0.28: the plans step and the plan's routes (addendum §39)
- A step of its own. The order of work is Lock, then the plans (proposed by Lock; adjust and OK each in Assets), then each place's backgrounds (one per direction, from the plan), then stills, then takes. The kickoff lists Plans: adjust and OK after Lock among its checkpoints.
- For the agent.
POST /api/projects/:id/storyboard/plan/places/:placeId/ok({ revision }) OKs a place's plan (and/ok/undotakes it back), free.project.places[i].planholdsnorth(degrees clockwise from up, where true north points),northBy,northWhy,revision,okAt,okBy,okRevisionandwarnedAt. A key frame's still carriesplan: the place, the area, the revision and the camera it was made with (x,yin that area's room,bearing,fov). After a plan change,POST …/storyboard/plan/remake/quoteprices the remakes of the stills and backgrounds it names (free),POST …/storyboard/plan/remake({ revision, placeId, what, items, maxCredits }) holds exactly that and queues one job each, andPOST …/storyboard/plan/keepkeeps them with a flag (project.planChanges), free.
Tarkovsky Director: the kickoff, "Decided for you" and run reports (0.18)
- Kickoff (project menu): the playbook's questions, asked once and kept with the project. Per area (story and script, looks, coverage and camera, edit): Ask me or Decide for me. The intent in a sentence or two (left empty, Script Coverage's logline stands in, marked as such), the budget in credits and the deadline, what to deliver (shot list, stills, takes, Rough Cut or export), checkpoints where a run pauses anyway, and how to reach you. Every step and the morning report read it; a run keeps the answers it started with. Viewers can read it; editors change it.
- Decided for you (project menu) is a view of History: every entry in Story's, the Shot List's and Edit's History that Tarkovsky Director (in an area you left to it) or an automatic step decided for you, with the question, the choice and where the choice came from (the script, the story's intent, Script Coverage, the genre, the easiest to change later…). Filter by place, source, who decided, or words. Change opens the place. Undo uses that History's own Undo: on the Shot List and in Edit it works from here; a Story change is undone in Story → History, where the script and the tags are.
- Run reports (the second tab) open each run's morning report in the playbook's order: Story flagged first when Script Coverage said Pass or its top notes called for serious dramatic changes in a run that decides the story (the film is finished as written; the proposed changes are a separate unlocked draft in Story → Drafts, never shot), then Decided for you, Blocked, What was made, the cost per stage against the budget, Agent or platform? with the two rates, and the decision log. The report as Markdown downloads the same report to post with the result.
- For the agent (API, editors write, any member reads):
POST /api/projects/:id/agent/runsstarts a run from the kickoff;…/runs/:run/decisionsappends the decision log (step, finding, do · ignore · flag, why, the "Decided for you" source, cost);…/problemsthe agent-or-platform rows;…/story-flagthe story gate;…/story-draft {revision, script}keeps the alternate draft;…/finishcloses the log with the cost by stage, what was made and what was blocked;GET …/report(or?format=md) reads it.
Tarkovsky Director: what changed in 0.20 (round 8), the kickoff comes first
- No paid step before the kickoff. While a project is in a Tarkovsky Director run, or its kickoff is marked required (Paid steps wait for this kickoff, outside a Tarkovsky Director run too), every paid step (stills, takes, voice, music, chats, Script Coverage, Propose coverage and direction, the review, the pre-take check, the Rough Cut) waits until the kickoff is answered. It says Answer the kickoff first. Until you do, only Brainstorm in the kickoff uses credits. (0.56), before anything is held. Quotes, Check script, the Breakdown, Lock and renders never wait.
- The new questions. Coverage density: Lean, Standard or Full, for the film, and per scene once the script is locked; Propose coverage opens on it, this scene's answer first. Protection and alternates: make them (they get stills and takes) or plan them only (taken off in 0.42: the amount of coverage covers it). The video model, chosen from the catalog, with what it carries (sound, references, a last frame) and its price a second. Remake limits: remake up to the limits or only flag; remakes a still, retakes a part, and the share of the budget all remakes may use. Answered by: who answered; the kickoff keeps when.
- The whole-film estimate. From the script, the density and the model, the film is costed to its export before anything paid starts: Story, Assets, the Shot List, stills, the pre-take check, takes with sound, remakes and retakes within the limits, the scripted music, the Rough Cut and the export (free). Each line says how it was reached. Over the budget, the kickoff shows three ways back, each with its new estimate: cut shots (protection and alternates planned only, then Lean in the costliest scenes), change the density, or raise the budget. The kickoff is saved once one of them fits.
- An unattended run takes the stored defaults (what was saved, the rest filled: Standard, planned protection, the project's video model, two still remakes and one retake within 20% of the budget). They are listed first in the run report under Decided for you. Over the budget, the defaults cut shots, then lower the density, and stop if the film still does not fit.
- For the agent:
GET /api/projects/:id/kickoff(answered, required, in a run);POST …/kickoff/estimate {kickoff}(free);POST …/kickoff {revision, kickoff, overBudget?}or{revision, defaults: true}. A paid route before the kickoff answers 409{ error, kickoff: true, agent }:erroris the filmmaker's line, andagent(0.56) says why it waits and how to wait for the form or take the defaults; the agent API and Tarkovsky Director's make give youagent.
Tarkovsky Director: what changed in 0.22, the report's cost from the ledger, corrections, the sign-in's end
- Cost by stage is what was charged (TAR-263). The run report's cost per stage now comes from the ledger: every reservation made for the project during the run, settled at its charge, under the stage it paid for (a take is Takes, a still is Stills or Assets, narration and music are Edit, a check by its own step). What is still held for calls running is shown apart. The decision log's costs stay beside it as a cross-check, and the report names the stages where the two differ.
- The catch rate counts only flags raised before the takes (playbook §7). A problem the takes, Edit or the review first named reached the cut uncaught, whatever the row says; the report counts those rows separately.
- A row can be corrected.
PATCH /api/projects/:id/agent/runs/:run/problems/:problem {flaggedBy?, acted?, easy?, owner?, source?, why}(editors, also after the finish). The owner is read again from the playbook's questions unless you name it; a row whose owner changes loses its approvals; every correction is kept on the row with who, when and why. A row can also say where the problem started (source: the still, the prompt, the plan or the model). - The sign-in's end. A sign-in lasts until a day without any request, and at most seven days after you signed in, whatever you do. A day before the seven days end, the studio says so with Sign in again, which saves your work first. For an agent:
GET /api/sessiongivessession.endsAtandsession.absoluteEndsAt; sign in again before a run that would pass it. - Saving from a tab that missed another change (an API write, a Lock, a still that landed) is now always Another session changed this project, with Reload and reapply. The reapply keeps every change you made that the other session did not also make, down to the field: a shot's description typed here while a still landed on that shot there both stay. Only a field changed both ways is put to you.
Tarkovsky Director: what changed in 0.22 (round 10 §6), the kickoff opens by itself
- It opens on its own. A project made from New project asks its kickoff first: the questions open as soon as the project is made, and paid steps wait for the answer (the kickoff is marked required). Closed with Later, it opens again the moment a paid step is tried, instead of an error. Nothing is reserved before it is answered. Projects made before 0.22, and projects made through the API, work as before.
- One question at a time. How much Tarkovsky Director decides for you, what the film is about, your budget and deadline, how many shots, the video model, remakes, the motion check, a split on a locked board, and what to deliver (since 0.42, only the questions for the areas you hand over; see below). Each question says what it means, and the whole-film estimate sits beside it and moves with every answer (since 0.50, beside the budget question only; see below). The last page lists your answers (click one to change it), the full estimate with the three ways back under the budget, and who answered.
- It can't be skipped silently. Use the defaults is an answer you give: the questions you haven't answered take the studio's defaults, and the kickoff records that you chose them, with your name (it is not listed under Decided for you; that list is for choices made without you). Later says that paid steps wait.
- Take resolution (new question after the video model, also in Project settings): the resolution every take is made at, chosen from the model's own list, each with its credits a second at that resolution. The default is the model's cheapest resolution that isn't below 720p (else its first). The whole-film estimate is priced at it, and a batch of takes or a take in Dailies uses it unless you choose another there; the quote says which.
- A split on a locked board (new question, also in Project settings): when a check wants to split a shot into two with a cut and the board is locked, Tarkovsky Director can ask me, unlock and split (logged in Decided for you, where Undo puts it back) or don't split, flag it. No answer reads as ask me.
- Every angle of a person (0.35, also in Project settings; since 0.42 Creating a character's additional angles): loving a person's look also makes its other angles, the total shown first (Make every angle when a look is loved, the default), or A cheap draft: only what I make. No answer reads as every angle. For the agent: the kickoff carries
everyAngle(false for the draft). - The answers live on the project. The long form (Kickoff in the project menu) shows every question on one page, to read or change later.
- For the agent: wait for the filmmaker's answer on the form, never in chat:
GET /api/projects/:id/kickoff?wait=60answers as soon asansweredis true (or after the wait).POST …/kickoff {revision, useDefaults: true, kickoff: {answeredBy}}is the filmmaker's own Use the defaults;{defaults: true}stays the unattended run's. The kickoff carriesansweredWith("answers" or "defaults") andsplitOnLockedBoard("ask", "unlock_and_split", "flag").
Tarkovsky Director: the kickoff in 0.42, only what you hand over
- Only what you hand over is asked. The first question is How much should Tarkovsky Director decide for you?, per area (story and script, looks, coverage and camera, edit): Ask me waits for your answer. Decide for me lets Tarkovsky Director choose, keeps a record, and can always be undone. After it come only the questions for the areas you hand over (Decide for me): the story for story and script; How many shots?, the video model, Preferred Resolution, the split on a locked board and the frame-by-frame check for coverage and camera (the check for the edit too); Creating a character's additional angles for looks; regenerating an image or a video for looks or coverage; and What to deliver, and when to pause once any area is handed over. How much to decide and your budget are always asked (since 0.44 the frame rate isn't a kickoff question: New project asks it). A question not asked takes the default, and Tarkovsky Director asks you when it comes up. The note at the top shows on the first question only.
- The story and the film's length. What is the film about, and how should it feel? If you'd like Tarkovsky Director to write the script, tell it the story in The story (optional): what happens, who the characters are, the themes, a line or a page. You can also write the story yourself or collaborate on the Story page. How long should the film be? (optional) is its own box, in minutes, with the page count beside them (one script page is about a minute, so a five-minute film is about five pages). Brainstorm under the story box is built and switched off until its words are approved; it doesn't show yet (0.44's two states are below).
- The whole-film estimate beside each question (since 0.50, beside the budget question only) moves with the film's length and the amount of coverage until the script is locked, and says Check again once the script is locked. Locked, it reads the locked scenes; with no length, the working script; with neither, only Script Coverage and the shot list's review.
- Protection and alternates is no longer asked: the amount of coverage covers it (an older kickoff that made them still counts them).
- Script is a deliverable: What to deliver → Script · the script only (and, since 0.44, Shot list · through the shot list). A Tarkovsky Director run then writes the script from the story (or finishes the one there), runs Check script, and ends once the script is ready for you to read: no Lock, no pictures, no export.
- Fewer words. How many shots?: choose the amount of proposed coverage you want; you can always add more shots yourself. The video model is listed once, in its menu; if your preferred model can't make a take (the reference images it takes, its length), Tarkovsky Director chooses one that can. Preferred Resolution is a choice and nothing more. When an image or video has an error (the character is inconsistent, looks the wrong direction, etc.) should Tarkovsky Director regenerate the image/video or only flag it?, with Regenerations for a still and Regenerations for a video. Would you like Tarkovsky Director to check each video take for errors, frame by frame?: (since 0.45) Yes, one vision call, about 0.25 credits per video, or No: You can check a video whenever you want; the cost shows before it runs. When Tarkovsky Director wants to split a shot and the board is locked.: Ask me, Unlock and split (listed under Decided for you; it can be undone) or Flag it. Tarkovsky Director will pause at these steps for your feedback and approval. Your answers uses the same words.
- For the agent: the kickoff carries
filmMinutes(1 to 180) anddeliver: "script";POST /api/projects/:id/kickoff/estimatereads the length until the script is locked (the estimate'sbasis:locked,length,scriptornone). Tarkovsky Director's finish needs no export when the kickoff delivers the script, once the script has a scene with words.
Tarkovsky Director: the kickoff in 0.44
- Always asked: how much to decide, and your budget. The kickoff no longer asks the frame rate: New project asks it, under the aspect ratio (24 by default), and Project settings change it. Your answers doesn't list it either. An older kickoff's rate stays the project's; saving the kickoff again never brings it back over the one set in New project or Project settings.
- The story box holds 4,000 characters. The story is its own field now, apart from the intent: the intent (Project settings, beside the director's note) and the logline keep their 2,000 characters and are no longer written by the story box. A kickoff answered before 0.44 opens with its story as it was. The story's new role: it goes to Tarkovsky Director to write a first draft of the script. Saving the kickoff sends nothing and charges nothing; since 0.48 the script is written in the run, from the synopsis (see the 0.48 section).
- Brainstorm on the story (still switched off, so it doesn't show yet) has two states. With the box empty: Say a few words (a place, a person, a feeling) and get a draft to edit. and Draft it. With words in the box: Brainstorm your idea. Say what to add, change, or leave it empty. and one button, Brainstorm; an empty message adds to the story. Its message box is empty, with no placeholder (0.46), as Assets' Brainstorm; the button says Brainstorming… while it works. Its price shows first (about 1.3 credits), and what comes back goes in the box for you to change. If it can't finish: Brainstorm couldn't finish. Nothing was charged.
- What to deliver: Script · the script only, Shot list · through the shot list, then stills, takes, a Rough Cut or the export.
- For the agent: the kickoff carries
story(up to 4,000 characters; an older kickoff's story is inintent, up to 2,000, which stays the intent line).POST /api/projects/:id/brainstorm/kickofftakesaction"draft"(withwords) or"brainstorm"(with thestoryand an optionalask); both answer 409switchedOffuntil it is switched on. AframeRatesent with a kickoff answer is still the project's before a cut.
Tarkovsky Director: the kickoff in 0.45
- The frame-by-frame check asks Yes or No. Would you like Tarkovsky Director to check each video take for errors, frame by frame? Yes checks each take as it lands (one vision call, about 0.25 credits per video). No: You can check a video whenever you want; the cost shows before it runs. (Check motion on a take in Dailies.) Your answers says Yes or No. A kickoff answered before keeps its answer.
- Only flag it means nothing is regenerated. When Only flag it is chosen, Regenerations for a still, Regenerations for a video and All remakes within, % of the budget all say 0, can't be changed, and are kept as 0: a still or video with an error is flagged for you, never remade by Tarkovsky Director. Choosing Remake it, up to the limits again brings the proposed limits back: a still up to 2 times, a video once, all remakes within 20% of the budget. The whole-film estimate counts no remakes for Only flag it, and its line says Nothing is regenerated. A still or video with an error is flagged for you. (0.46). A remake you make yourself is never blocked, and under Only flag it it no longer says "over the limit". The Shot List's Remakes works the same way (Flag only).
- For the agent: an answer with
remakeLimits.mode: "flag"is kept asstillRemakes: 0, takeRetakes: 0, budgetShare: 0, and the stored defaults are Only flag it at 0 (they were 2, 1 and 20%).
Tarkovsky Director: the kickoff in 0.46
- The remakes answer and the Shot List's Remakes are one setting. Changing it in the kickoff changes the Shot List's, and changing it on the Shot List changes the kickoff's. Choosing Remake it, up to the limits brings back 2, 1 and 20% in both. With neither set, both say Only flag it at 0. An older project where the two differ shows the Shot List's in both: it was the later change.
- For the agent:
POST /kickoffwithoutremakeLimitskeeps the project's setting, the Shot List's, rather than the stored default. A project with nothing set refuses an automatic or agent remake (409,remakeLimit.code: "flag_only") before anything is held.GET /remake-limitsanswerssource: "default"with 0, 0 and 0.
Tarkovsky Director: the kickoff's story box in 0.48
- One box for the story. What is the film about, and how should it feel? The story (optional) takes your notes, as little or as much as you like, up to 4,000 characters (N / 4,000 in the box's foot once there's text; since 0.50, for notes only). The box's foot: Brainstorm on the left, the count, and Write the synopsis on the right once there's text. How long should the film be? is unchanged.
- Write the synopsis turns your notes into a synopsis Tarkovsky Director can write the script from, sized to the film's length, and replaces the box's text with it; a sentence under the box says what it did. It's free. You can edit the synopsis like any text.
- The doc icon (top right inside the box, once there's text) opens the text larger to read: the film's title, then The synopsis once Write the synopsis has written it (otherwise The story), with Edit (edit it right there; the button becomes Done), Brainstorm and ×. Esc or a click outside closes it. Edit also opens Brainstorm beside it (below it on a phone): Change it on the left yourself, or tell me here. A change you ask for there lands straight in the text you're editing. Done or × puts Brainstorm back in its column.
- Brainstorm is a conversation about the story, in the column on the right of question 2 (below the box on a narrow screen; under the question in the long form, Kickoff in the project menu). Its first line: Let's talk the story over. Ask anything, try ideas, and I'll keep a draft beside this window up to date. Ask a question and you get an answer, and the draft stays as it is. Ask for a change and the draft is reworked all the way through, on The draft, a window beside the conversation (above it on a phone), which starts as the story box's text. Change the draft there yourself too (Change it here, or in the chat.); what's on it, your edits included, is what the next message works on. Use this puts the draft in the story box and closes it; the conversation stays open, and the next change opens the draft again. × on the draft closes it without using it (since 0.50 the draft is kept, waiting, until Use this; see below). After the first answer, the message box suggests a next step, grayed (or Change the…); Send with the box empty sends the suggestion. The mic beside it dictates where your browser can (it doesn't show where it can't).
- Its price: About 1–3 credits a message (since 0.51; it said About 1 credit a message), and each answer shows what it cost. The most a message can cost is held first and only what it used is charged; an answer that doesn't come back costs nothing (Brainstorm couldn't finish. Nothing was charged.). One message at a time. The conversation lasts while the kickoff is open; what you keep is what's in the story box when you save the kickoff (and, since 0.50, a draft you haven't used yet).
- A commercial kept the story box as before, without Brainstorm and Write the synopsis, until 0.49 (below).
- For the agent:
POST /api/projects/:id/brainstorm/kickoff/quote(free) andPOST /api/projects/:id/brainstorm/kickofftake{ notes, draft, conversation: [{ from: "filmmaker" | "you", text }], message, minutes?, maxCredits }and answer{ reply, draft, next, chargedCredits, releasedCredits }:draftis the whole rewritten draft, ornullwhen the story didn't change. 0.42'saction,words,storyandaskare gone. The kickoff carriesstoryKind("notes", or"synopsis"once Write the synopsis has filled the box, through later edits until the box is emptied).
Tarkovsky Director: the synopsis and the script from the story (0.48)
- Write the synopsis is free. It turns the notes in the story box into a synopsis Tarkovsky Director can write the script from, sized to the film's length (about a page a minute of film: a paragraph for a minute or two, about a page for fifteen minutes, several pages for a feature). Nothing is charged; the studio pays for it. It reads the film's title, type and length (5 minutes when the length box is empty) and your notes. Pressing it again with the same notes, length, title and type gives the same synopsis back without writing it again. You can write up to 20 synopses an hour (That's enough for this hour: try again later.). If one doesn't come: No synopsis came this time. Please try again. With nothing in the box: Write some notes in the box first.
- The script is written in the run, from the synopsis. When you hand the story over (Decide for me for the story and script, or What to deliver → Script), a Tarkovsky Director run writes the script from the kickoff's story into the project's script, then carries on as it does from a written script (Check script, and so on). If the story is still your notes (you didn't press Write the synopsis), the run writes the synopsis first, with the same words, and keeps it as the kickoff's story: the kickoff then shows The synopsis. Your notes are kept with it.
- Since 0.56 the studio starts it without a run. When Story opens with an empty script, Decide for me for the story and script, and a synopsis kept in the kickoff (a concept on a commercial), the studio writes the script on its own, once a project: Writing the script from your synopsis, so you have a first draft to work on: about 80 credits. A feature shows the part being written (Part 2 of 4) and Stop, which lets that part land and holds nothing more; after a Stop, or a part that didn't land, it never starts again by itself, and Write the rest (with its price) carries on. Without enough credits it doesn't start and says so. For the agent:
GET /api/projects/:id/kickoff/script/autoreads it; its start (POST) and Stop (DELETE) are the page's, refused to agents, who write the script withPOST /kickoff/scriptas before. - Never over a script. If the project has any words in its script, yours or a locked draft's, the run never writes one. With Ask me for the story (and any deliverable but the script), the script stays yours to write or import.
- A feature is written a part at a time. Up to 30 minutes, the script is written in one go. Longer, in four parts: act one, act two up to its midpoint, act two from its midpoint, act three, each carrying on from the pages before it (eight parts, in runs of scenes, past two hours). Each part is in the project's script as soon as it's written, and a run that delivers the script ends only once the last part is in.
- What it costs. The script, and the synopsis the run writes first, are paid from the run's budget: each call shows its price first, holds its most, and is charged what it used; a call that can't finish costs nothing. The whole-film estimate has a Script line while the script is still to be written: Tarkovsky Director writes the synopsis from your notes, then the script from it: about 5 pages. (or …the script from your synopsis…, with the number of parts for a feature).
- Commercials kept today's flow in 0.48; since 0.49 a commercial has Write the concept and the script from it (below).
- For the agent:
POST /api/projects/:id/kickoff/synopsis {notes, minutes?}→{said, synopsis, model}(editors; free; 409 with no notes). It is the filmmaker's button: Tarkovsky Director and outside agents are refused it.POST /api/projects/:id/kickoff/script/quote(free) → the next call's steps andmaximumCredits;POST /api/projects/:id/kickoff/script {maxCredits}(paid) writes the synopsis when the story is notes, then the next part:{part, done, next?, …}; call it untildone. While it's due, a save of a script into an empty one by an agent or by Tarkovsky Director is refused (409scriptWrite). The kickoff carriesstoryKind("notes"or"synopsis"; absent reads as notes),storyNotes, andscriptWrite {parts, done};storyholds notes up to 4,000 characters or a synopsis up to 40,000. A story saved withoutstoryKindthat changed reads as notes.
Tarkovsky Director: a commercial's kickoff, from an idea (0.49)
- Question 2 on a commercial. What is the commercial about, and how should it feel? Under it: If you'd like Tarkovsky Director to write the script, tell it the idea here: the product, who it's for, what the spot should say. A line or a page, as simple or as detailed as you like. The client, the mandatories and the brand rules go on the Brief page. It's asked when the story and script are handed over, as on a film.
- The idea (optional) is one box, up to 4,000 characters (N / 4,000 in its foot once there's text). Its foot: Brainstorm on the left, the count, and Write the concept on the right once there's text. It works the way a film's story box does (0.48), with the commercial's words.
- Write the concept turns the idea into the spot's concept, sized to the longest length chosen, with a line on what each shorter cut keeps, and replaces the box's text with it. A sentence under the box says what it wrote. It's free. It also proposes the Brief page's empty fields (the product, who it's for, the one message): when you save the kickoff they're filled in on the Brief page, where you can change them like any field. A field that already has words is never touched, and the client, the mandatories and the brand rules always come from the Brief page.
- How long should the spot be? (optional) 6, 15, 30 and 60 seconds, as chips; choose more than one for shorter cuts of the same spot. They are the Brief page's lengths: the chips start on the Brief page's (30 when it has none), and saving the kickoff writes them back. Write the concept and Brainstorm read the chips as they stand, before you save. On a commercial they take the place of the film's How long should the film be?.
- Brainstorm · The idea is a conversation about the spot, in the column on the right: Let's talk the spot over. Ask anything, try ideas, and I'll keep a draft beside this window up to date. A question gets an answer and leaves the draft alone; a change reworks the draft on The draft beside it, and Use this puts it in the box. About 1 credit a message, as on a film.
- The doc icon opens the text larger to read: the spot's title, then The concept once Write the concept has written it (otherwise The idea), with Edit, Brainstorm and ×, as on a film.
- When you hand the idea over without Write the concept, the run writes the concept first, then the script, as a film's run does with its synopsis.
- Your answers (the kickoff's last page) shows the idea and the lengths: … · 30, 15 and 6 seconds.
- For the agent:
POST /api/projects/:id/kickoff/concept { idea, lengths? }(free) answers{ said, concept, brief, model };briefis{ product, audience, keyMessage }(each a proposal ornull) ornull.POST …/kickofftakesspotLengths(each 6, 15, 30 or 60; the Brief page's lengths) andbriefFill({ product?, audience?, keyMessage? }, filling only empty fields) beside the kickoff, on a commercial. The kickoff's Brainstorm routes takelengthsin place ofminuteson a commercial.
Tarkovsky Director: Write the concept, and a commercial's script from it (0.49)
- Write the concept is free. On a commercial's kickoff it turns the idea in the box into the concept Tarkovsky Director writes the script from, sized to the spot's longest length, with a line on what each shorter cut keeps. Nothing is charged; the studio pays for it, as for Write the synopsis. It reads the film's title, the spot's lengths, the brief as far as it's filled in, the brand rules and your idea. Pressing it again with the same idea and the same brief, brand rules, lengths and title gives the same concept back without writing it again. You can write up to 20 concepts an hour and 60 a day (That's enough for this hour: try again later.), apart from your synopses. If one doesn't come: No concept came this time. Nothing was charged: try again. With nothing in the box: Write the idea in the box first.
- The brief from the concept. The concept also proposes the brief's fields that are still empty: the product, who it's for and the one message. The studio fills those empty fields on the Brief page when the kickoff is saved; you can change them there like any field. A field already filled is never touched, and the client, the mandatories and the brand rules always come from the Brief page.
- The spot's lengths are the Brief page's. The kickoff's chips (6, 15, 30 and 60 seconds) and the Brief page's lengths are one list: saving the kickoff saves the chips as the brief's lengths, longest first. With none chosen, the calls use 30 seconds.
- Brainstorm on a commercial works as on a film, with the commercial's own prompt: it reads the film's title, the spot's lengths, the brief as far as it's filled in, the brand rules, the idea, the draft and the conversation so far. About 1 credit a message, as on a film.
- The script is written in the run, from the concept. When you hand the idea over (Decide for me for the story and script, or What to deliver → Script), a Tarkovsky Director run writes the commercial's script into the Brief page's picture and sound table, so Build shot list works as before. Each row has its seconds (the shot's length; the rows add up to the spot's length), and each shorter length gets its cut: the rows it keeps. If the kickoff still holds your idea (you didn't press Write the concept), the run writes the concept first and keeps it as the kickoff's story (the kickoff then shows The concept; your idea is kept with it), and fills the brief's empty fields from it. A spot is written in one call.
- Never over a table with words. If the Brief page's table has any row with words, the run never writes one. A run that delivers the script ends once the table has its rows.
- What it costs. The script, and the concept the run writes first, are paid from the run's budget: each call shows its price first, holds its most, and is charged what it used; a call that can't finish costs nothing. The whole-film estimate's Script line on a commercial reads Tarkovsky Director writes the concept from your idea, then the script from it: the 30-second spot as the Brief page's picture and sound table, in one call each. (or …writes the script from your concept…, in one call.).
- A kickoff that delivers the script is priced as the script alone. Its whole-film estimate is the Script line only (the film's other lines are left out), and Tarkovsky Director's panel says what is left is the script and prices the script's own calls (every part of a feature), not the film's stills, takes or edit.
- Draft from brief on the Brief page writes with the same prompt as the run: from the concept when the kickoff has one, from the brief otherwise, with the table as it stands and the Draft panel's direction. Its rows come with their seconds and its shorter cuts, as proposals: nothing is added to the table without a click.
- Longer scripts land. A part of a feature's script that thought for longer than usual could stop at the output limit (The answer was cut off at the output limit before it finished., nothing charged). Every script, synopsis and concept call now has room for that thinking; the most it can hold is a little higher, and it is still charged only what it used.
- For the agent:
POST /api/projects/:id/kickoff/concept {idea, lengths?}→{said, concept, brief: {product, audience, keyMessage} | null, model}(editors; free; 409 with no idea or on a film). It is the filmmaker's button: Tarkovsky Director and outside agents are refused it. The kickoff's save (POST /api/projects/:id/kickoff) also takesspotLengths(each 6, 15, 30 or 60) andbriefFill({product?, audience?, keyMessage?}, empty fields only) on a commercial, and ignores both on a film. On a commercial,POST /api/projects/:id/kickoff/script/quote(free) andPOST /api/projects/:id/kickoff/script {maxCredits}(paid) write the concept when the kickoff holds the idea, then the table:{concept?, part: {n: 1, of: 1}, done: true, rows, cuts, chargedCredits, releasedCredits}. While it's due, a save that puts words into the empty table by an agent or by Tarkovsky Director is refused (409{scriptWrite: true}). The brief's rows carryseconds; the brief carriescuts: [{seconds, rowIds}]. The kickoff's estimate answersthrough: "script" | "export".
Tarkovsky Director: changes are highlighted while editing (0.49, film and commercial)
- In the reading window, after Edit, every word that has changed since you pressed Edit is highlighted (amber, behind the text), whether you typed it or Brainstorm made it beside the window. Removed words aren't shown, and neither are spaces or line breaks on their own; a word you moved is highlighted where it landed.
- The highlights go when Done or × ends the editing. Pressing Edit again starts from the text as it is then.
- A very long text written as one paragraph, changed in two places far apart, is highlighted from the first change to the last; with paragraphs, each change is highlighted on its own.
Tarkovsky Director: the kickoff and its story box in 0.50
- The whole-film estimate sits beside the budget question only (Your budget, and when it is due), with its Script line there while the script is still to be written. The other questions have nothing beside them and take the window's width; question 2 keeps Brainstorm's column. In the long form (Kickoff in the project menu) the estimate is under the budget. Your answers, the last page, still shows the whole estimate and the three ways back under the budget.
- The deadline is asked only when you hand something over. With at least one area on Decide for me, the budget question asks for the budget and the deadline. With every area on Ask me it's Your budget alone, and no deadline is kept: with nothing handed over, there's nothing to schedule.
- Brainstorm keeps its draft, as Brainstorm on Assets does. A draft Brainstorm wrote waits until you press Use this: × on the draft, or on Brainstorm, closes it without losing it, and so does closing the kickoff or leaving the page. While one is waiting, Brainstorm shows a dot (Your brainstorm is waiting, as you left it), and opening Brainstorm opens the draft again where you left it. Only Use this puts it in the story box. Your next message works on the waiting draft. The conversation itself still ends when the kickoff closes. Films and commercials alike. The draft is yours: kept for you, per project, not for the others on it.
- A long synopsis. A synopsis (or a commercial's concept) shows in full in the story box, which scrolls; the N / 4,000 count shows for notes (an idea) only. The reading window scrolls too.
- Download. The reading window has a download button beside × (Download). It saves the text as it stands, your unsaved edits included, as a plain text file named after the film: Emina's Crossing - The synopsis.txt (or - The story.txt; on a commercial - The idea.txt or - The concept.txt).
- New words. Write the synopsis: No synopsis came this time. Please try again., The synopsis is being written., and, where it can't run, Write the synopsis isn't available right now. Please try again later. (That's enough for this hour: try again later. stays.) Notes over the limit, in the box or at the kickoff's save: The notes hold 4,000 characters. For Tarkovsky Director and agents, the script's The script is being written.
- For the agent: the kickoff's save drops
deadlinewhen no area inmodesis"autonomous". The kept draft is a row of Brainstorm's own (GET/PUT /api/projects/:id/brainstorm/kept, itemkickoff:story, route"text", the draft as one part;POST …/brainstorm/kept/clearat Use this): per person, never the project's data.
Tarkovsky Director: the kickoff and its story box in 0.51
- Brainstorm works on a long synopsis or concept. A synopsis (or a commercial's concept) longer than 4,000 characters no longer grays out Brainstorm: it opens from the box, from the reading window, and beside the text in Edit, on a film and on a commercial alike. It reads the whole text, up to the box's 40,000 characters, and its draft can be as long as the text it works on: it never cuts a long synopsis down and rewrites it shorter. Use this puts a draft longer than 4,000 characters in the box as a synopsis (a concept on a commercial), whole; a waiting draft is kept whole too.
- Its price, before you send. Under 4,000 characters the line beside Send reads About 1–3 credits a message. Over 4,000 a message costs more, so the line shows that message's own price, from a free estimate: This message: about 6.2 credits, up to 14 (Pricing this message… while it's read; hover it for why). Send waits for it. The most it shows is the most that's held when you send; only what the message uses is charged, and an answer that doesn't come back costs nothing. If the price went up in the meantime, nothing is sent: The price changed: it's beside Send now. Send again to go ahead.
- No length chosen keeps the Brief page's. On a commercial, saving the kickoff with none of How long should the spot be? chosen (the short form, the long form, or Use the defaults) leaves the Brief page's lengths as they are; before 0.51 it emptied them. Choosing lengths still writes them to the Brief page.
- New words. Your answers with no idea: No idea yet: Tarkovsky Director will work from the Brief page. A run whose idea you changed while it wrote the concept: You changed the idea while the concept was being written. Your new idea has been kept; the concept from the old one is below, but has not been saved.
- Draft from brief follows your direction. The commercial's script prompt now says so: when you give a direction in the Draft panel's box, the draft follows it within the brief, the brand rules and the mandatories.
- For the agent:
POST /api/projects/:id/brainstorm/kickoff/quoteandPOST …/brainstorm/kickofftakenotesanddraftup to 40,000 characters each (was 8,000) andmaxCreditsup to 600 (was 100);draftcomes back whole, up to 40,000. A quote with an emptymessagegives the most for the longest message (2,000 characters), the expected figure for a short one. On a commercial, the kickoff's save leaves the brief's lengths whenspotLengthsis[]or left out.
Tarkovsky Director: the kickoff's words in 0.55
- Zach's newer words. The remakes question is When a still or take has an error (a character looks different, faces the wrong way, and so on), should the studio remake it or only flag it?, with Remakes for a still, Remakes for a take and Remakes may use up to this % of the budget; the split question is When Tarkovsky Director wants to split a shot but the Storyboard is locked, what should it do?, with Ask me, Unlock and split or Flag it (a saved answer reads as before), and a kickoff no one answered says it was answered by the defaults (no one was there to answer).
The agent path: assistants and agents at /mcp (0.47)
- What it is. Tarkovsky is an MCP server at
https://tarkovsky.ai/mcp. An assistant or agent you connect works for you in your account. That can be Claude Code, Cursor, ChatGPT, Claude or your own script. It reads a project's script, shots, storyboard, takes and cuts. It makes a still or a take for a planned shot, and it renders the cut. Everything goes through the same quotes, holds and gates as the page and Tarkovsky Director: the kickoff, the project's budget, the remake limits and the Terms. - A budget per connection. Each connection (an agent key, or a sign-in from an assistant) has its own budget in credits: 200 unless you set another. A paid call whose ceiling would pass the budget is refused before anything is held. The connection shows what it has spent.
- Connecting with an agent key. Signed in,
POST /api/account/agent-linkswith{"name": "…", "budgetCredits": 50}makes a key. Its answer shows the key (tk_…) once; keep it. Send it asAuthorization: Bearer tk_…to/mcp, for exampleclaude mcp add --transport http tarkovsky https://tarkovsky.ai/mcp --header "Authorization: Bearer tk_…". GET /api/account/agent-linkslists your connections and what each spent.PUT /api/account/agent-links/:idchanges a name or budget.DELETE /api/account/agent-links/:idrevokes one at once.- An account keeps at most 20 keys. A page in the studio for them comes later.
- Signing in from ChatGPT and Claude waits on the studio's sign-in setup and the directories' review. Until then, connect with an agent key.
- Also: REST at
/agent/v1, described at/agent/v1/openapi.json. Send anIdempotency-Keyheader so a repeated request is answered once. - Cutting (0.59). An assistant can look at a take (its contact sheets, where someone speaks and, once the take is transcribed, every word with its time) and write the cut of a scene or of the whole film from your loved takes. The studio checks it as it checks the Rough Cut: only loved takes, and on a shot with lines no cut while someone is speaking (between two words is fine once the take is transcribed). It's free. It makes a new version of your edit, so the one before stays in Sequences, and Edit's History can undo it.
- The whole studio (0.62). A connection can do what the studio's pages do on your projects, one request at a time, with the same checks and prices.
- An agent with a shell calls the studio's own routes at
https://tarkovsky.ai/agent/v1/api/…with the same key, and downloads a file at/agent/v1/media/<id>. - An assistant does the same through its
studio_requesttool. /agent/v1/api.mdlists every route open to a connection, and what each one costs.- Every credit a connection holds counts against its budget before anything is held.
- Deleting, buying credits, sharing, publishing to the Gallery, confirming the rights to a picture and answering a flag stay yours, in the studio.
GET /api/account/agent-linksalso shows, for each connection, which way it called in the last 30 days.- The API, in full (0.62.2). The API page documents every tool with its inputs, every route open to an agent with what it costs, the scopes, budgets, errors and examples. Agents read the same document as Markdown at /help/api.md. Both are written from the code with every release.
Tarkovsky Director: what changed in 0.22, stills in a run are checked before they are loved
- Every still is checked as it lands. A still's quote (
POST /api/images/quotewith the still'starget) includes its still check:stillCheckgives its price,imagethe still's own. Send the total asmaxCreditsand the worker checks the still when it lands. A clean first still is loved; one with findings lands unloved, with its notes (GET /api/projects/:id/pretake/still?mediaId=);GET …/pretake/still/stateslists every frame still's state (checking, clean, minor, findings, didnt_run, unchecked, loved_unchecked). In a run, a still is never loved unchecked as it lands, whatever its price. - Tarkovsky Director loves only checked stills. In a run (the run's header, or a run you started), a save that loves a frame still with no still check is refused: 409
{ error, stillCheck: { mediaId, keyframeId, frame, check, quote } }, naming the check and its price. Check it (POST …/pretake/still), then love it; with findings, love it anyway (POST …/pretake/still/decision {decision: "love_anyway"}) or remake it. Love without the check is your choice: in a run it is refused the same way; you (outside the run) can still choose it, and the still's Info panel says Loved without the check. - One check per still. A second
POST …/pretake/stillfor a still being checked (as it landed, or by another request) waits for that check and returns its result withjoined: true: one call, one charge. - A run ends with every frame's still loved, or says why.
POST …/agent/runs/:run/finishis refused (409{ error, unloved: [{ keyframeId, frame, stills, checks }] }) while a key frame has stills and none loved (a shot switched off is left out). Love one, or say why for each:unloved: [{ keyframeId, why }], or ablockedrow that names the frame. The report lists them under Blocked as 1C · FIRST (1 of 1): no loved still.
Tarkovsky Director: what changed in 0.23 (addendum §28), a must-fix flag is never waved through
- In a Tarkovsky Director run a must-fix flag closes only with a fix or your answer. That is every check's flag marked must-fix: the pre-take check's notes, the set-up and writing checks', the still check's, the motion check's findings and the take's text check (writing in the picture is always must-fix). Tarkovsky Director fixes it (remakes the still, retakes the part, or changes the plan and checks again), or stops and asks. It is never accepted to meet the deadline.
- The studio refuses Tarkovsky Director, as it refuses an unchecked Love: an outcome of ignored, Love anyway, a still or take loved over the flag, a take of a part whose stills carry one, and ignore in the decision log all answer 409
{ error, mustFix: { flags, fix, ask, deadline } }, naming each flag. Nothing is reserved or written. - A minor flag may still be accepted, with its reason (
whyon the outcome, Love anyway or the decision); without one the answer is 400. The report lists each with its reason. - Since 0.29 it stays accepted in the run's flags, with its reason, whatever comes after: a later check that no longer finds it, or its still set aside. Before, it could read fixed.
- You keep your choices. Outside a run (or from another member's account) ignoring a note, Love anyway, loving a take and making a take work as before. Love anyway answers every flag on that still, one Tarkovsky Director marked for a remake included, so its takes can go ahead. It says who loved it anyway and when, on the run's report, Tarkovsky Director's panel and the still's Info. A later check of the same still, in a Tarkovsky Director run or not, keeps those flags closed, and a Remake by an agent or a Tarkovsky Director run doesn't reopen them; a new still on the frame (a remake, an edit, an upload, or going back to an older one) has flags of its own. To answer a flag Tarkovsky Director asked about, open Decided for you → the run's report → Flags → Answer: Accept closes it; Fix it sends it back. While a run is open, the account that started it counts as Tarkovsky Director.
- If time runs out, the deadline moves, and the report says so: a run that finishes after its deadline moved it, and Tarkovsky Director records each move with why (
POST …/agent/runs/:run/deadline { to, why }). - The report's Flags section (after Blocked) counts the must-fix flags fixed, asked (answered by you among them) and left, lists the minor ones accepted with their reasons, and gives the deadline.
GET …/agent/runs/:run/flagslists every flag of the run with its id and where it stands; the decision log names a flag by that id (findingId) or gives a finding its ownseverity. - Which flags stand (0.26.2, TAR-318). One rule decides, for Tarkovsky Director's loves, its takes and the run's flags alike. A still's note stands while the frame still has that still and no later check has judged the same still again.
- A later check on the same still ends a note when it judges the same things: the pre-take check (or its re-check) ends any note; the still check ends its own notes and the code's; the free code checks end only the code's. A clean check counts, so changing the plan and checking the same still again, with nothing found, ends its notes.
- A check of another still doesn't: a remake that is not loved, checked or not, leaves the loved still's notes standing, because the loved still is what a take sends.
- Marking a note acted doesn't fix it on its own: it is fixed once its still is replaced or checked again clean. Until then the flags list shows it left, as the guards read it.
- Fixed by replacing the still is fixed while that still stays aside (0.28, TAR-337). The flags list names the still a note was raised on (
appliesTo). When the note is fixed only because another still is loved on its frame, the list saysstandsIfLoved: love that still again and the note stands again. In a run Tarkovsky Director's love of it is refused (409), and the refusal says why the list showed it fixed. A note fixed by a clean check of its own still has nostandsIfLoved: that still can be loved again.
Tarkovsky Director (0.27)
- What it is. The Tarkovsky Director button sits right of the stage tabs, after Export. It lets an AI model direct and produce the film from wherever it stands, until it is finished or you stop it. Tarkovsky Director works on the studio's own server, so you can close the tab. It acts as you, on this project only, through the same routes, checks, guards and prices as the page. Its stills and takes are built by the same code the page uses, from your plan: it never writes an image or video prompt itself. Every call it makes is its run's, so the run's rules apply: no unchecked love, and a must-fix flag is never accepted.
- Before a run. The panel shows what is left (scenes without a shot list, frames with no loved still, parts with no loved take, the asset images the plan needs, the cut and the render). It also shows the estimate to finish, low to high, split into studio work and directing (the model's own turns). Choose the model from the ones Admin offers and set a budget. The budget is the high estimate by default, and never more than your wallet or the project's budget left. Start opens a Tarkovsky Director run, filling in the kickoff with every area on "decide for me" when the project has none.
- What it costs. Directing is priced like every paid call: the model's token cost × 1.5, reserved before each turn and settled on the provider's receipt. The studio work it orders is priced as usual. The budget caps both together. A paid call that would pass it is refused, and when the next turn would pass it, Tarkovsky Director stops by itself, the way Stop does.
- Directing is most of the cost. A turn reads the project and its own notes, so it costs far more than a chat reply. The panel says what a typical turn costs on each model the menu offers, and which is the least expensive. The studio work costs the same whichever model directs. On a short film, expect the directing to be about as much as the stills and takes together, or more.
- While it works. The panel shows the model, the step it is on, its spend against the budget (directing, studio, total) and its latest log lines: turns, studio calls, makes, decisions and its eyes-pass findings. You can keep working in the studio. Tarkovsky Director reads the project fresh each turn, and a change of yours is never overwritten. Any member can watch; editors start, stop and resume.
- Stop. Stop finishes the current step. Tarkovsky Director starts no new paid work, lets the stills and takes already running land, and writes its account: what it did, what it would do next, what is left. The report goes into the log and into a PDF (Download the report (PDF)). Resume carries on from there, with the same conversation and an optional new budget. Without one, Resume uses what was left of the old budget.
- The report. It covers why the run ended, Tarkovsky Director's own account, what it made, its decisions, the eyes pass (each finding through a character's eyes or the viewer's), the flags (TAR-318's rule decides which still stand), the problems it found and whose they are (agent or platform), the spend against the budget, and what is left. The full log is an appendix.
- What it may change. Tarkovsky Director works within your kickoff. It may change the kickoff's production choices when the film needs them, and logs why: the coverage density (per scene too), protection, the video model and its resolution, the motion check, splits on a locked board, the frame rate. Your budget, deadline, modes, checkpoints, what to deliver, the intent and the remake limits stay yours: the studio refuses Tarkovsky Director a change to any of them. It makes every still, take, asset image and set pack through the same builders the page uses, never through a route of its own.
- Directions. Tarkovsky Director follows the agent playbook (v0.10): the steps in order, every finding decided, a must-fix flag fixed or asked about, and the eyes pass on every scene. The eyes pass runs three times, on the shot list, the loved stills and the takes. It first goes through each character's eyes (where they are going, what they look at, what they know), then the viewer's (causes before effects, what the viewer knows holds). The script and the characters' intentions win over the pictures: a wrong still is remade, never explained away by changing the plan.
- Who can use it. Everyone signed in with a verified email, while Admin's switch is on and set to Everyone (below). Admin can narrow it to admins, or admins and invited testers.
- The report's dates (0.50, Zach's Q2): the PDF writes each day as October 25, 2026 (UTC), as the Cover page writes a date; until 0.50 it wrote 25 October 2026. The Terms of Service's version line writes its date the same way (Version 1.0 · October 12, 2026); no word of the Terms changed.
Tarkovsky Director, shot by shot (0.69; off until Admin turns it on)
- What changes. A run that works shot by shot makes each shot right before the next one in its scene: the scene's master first, then its main shots in the script's order, then the protection and alternate shots, each once the main shot it is made from is locked. A master that starts on the shot before it goes after that shot, and a shot that starts on another scene's shot waits for that shot. For each shot it makes its frames (the first, any key frames, and the last), and the still check reads each one. Only once every frame passes does it make the take, and the Critic on Dailies watches it. Once both pass, the shot is locked and is the reference for the next shot. Scenes still go side by side.
- Tries. At each step a shot gets a few tries: stills at its frames, then takes at its daily. After those, a few whole-shot loops, each with the shot's words rewritten or another model. Admin › Tarkovsky Director sets how many. In a shot-by-shot run these replace the project's remake limits for the run's own makes; a person's remakes and an outside agent's keep them. A run keeps the numbers it started with. Take the best version doesn't apply in such a run: a must-fix note is always the loop's to fix. A check that can't run (the still check or the Critic on Dailies) counts as a try, never as a pass, and the run's report lists it. A run that makes no takes (its kickoff delivers the stills) locks a shot once its frames pass.
- Credits. Each shot has an allowance: the run's credits left divided by its shots left, when the shot starts. A hard shot may spend up to a few times that (Admin sets how many), drawn from the rest of the run. Its still checks and the Critic's watches of its takes count too, and so does any check Tarkovsky Director asks for itself. Past that, nothing more is made or watched for the shot. When three shots in a row are parked on their credits, the budget can't pay for a shot at that rate: the run stops as out of budget and emails you at once, and you can give it a larger budget and resume it. A shot whose next try the run's budget can't pay is parked too.
- Parked shots. A shot that runs out of tries or credits is parked. You get one flag for it, a line in Tarkovsky Director's panel and an email to whoever started the run (one an hour at most: the shots parked in between come in one email, sent at once when the run stops or ends). The run goes on with the other shots; only the shots that start from a parked shot, directly or through the shots between, wait. Change the shot, then press Try again in the panel: it goes back in the queue with fresh tries. A stopped run makes it when you resume it.
- No waiting on a note. A must-fix note is the run's to fix with its tries, so the run never stops to ask you about one. Once only parked shots, and the shots that wait on them, are left, it waits for your Try again.
- How a run is chosen. A run reads this once, when it starts, and keeps it through Resume. Admin's switch turns it on for every new run, and Admin can set one project on or off whatever the switch says.
QUESTIONS · TARKOVSKY DIRECTOR
Questions, answered from the guide.
- What does a Tarkovsky Director run cost?
What it costs. Directing is priced like every paid call: the model's token cost × 1.5, reserved before each turn and settled on the provider's receipt. The studio work it orders is priced as usual. The budget caps both together. A paid call that would pass it is refused, and when the next turn would pass it, Tarkovsky Director stops by itself, the way Stop does.
- Can I stop a run?
Stop. Stop finishes the current step. Tarkovsky Director starts no new paid work, lets the stills and takes already running land, and writes its account: what it did, what it would do next, what is left. The report goes into the log and into a PDF (Download the report (PDF)). Resume carries on from there, with the same conversation and an optional new budget. Without one, Resume uses what was left of the old budget.
- Can Tarkovsky Director wave a must-fix flag through?
In a Tarkovsky Director run a must-fix flag closes only with a fix or your answer. That is every check's flag marked must-fix: the pre-take check's notes, the set-up and writing checks', the still check's, the motion check's findings and the take's text check (writing in the picture is always must-fix). Tarkovsky Director fixes it (remakes the still, retakes the part, or changes the plan and checks again), or stops and asks. It is never accepted to meet the deadline.
- What is the kickoff?
Kickoff (project menu): the playbook's questions, asked once and kept with the project. Per area (story and script, looks, coverage and camera, edit): Ask me or Decide for me. The intent in a sentence or two (left empty, Script Coverage's logline stands in, marked as such), the budget in credits and the deadline, what to deliver (shot list, stills, takes, Rough Cut or export), checkpoints where a run pauses anyway, and how to reach you. Every step and the morning report read it; a run keeps the answers it started with. Viewers can read it; editors change it.