A pipeline skill, sanitized

The blocking pipeline

This is the master skill, with its six sub-skills, for scenes whose casts are too large for a keyframe per shot. It replaced an earlier flow of per-shot elevations and independent generations that never agreed on geometry. The idea is cheap: pin positions on a floor plan and one approved photoreal master, then derive every angle from that master inside one multi-cut clip, and check everything checkable on a still before any video is bought. Every rule below names the failure that taught it, including when to stop regenerating and hand the decision back to the creator. Project, scene, character and colleague names, paths and commands are removed. Download it as SKILL.md.

---
name: blocking-pipeline
description: Master skill for the blocking pipeline, the flow for scenes whose casts are too large for a keyframe-per-shot approach. Script → shot list with a verbatim BLOCKING block → an oblique location plate and three-panel model cards → a code-drawn floor plan with every camera stamped → ONE creator-approved photoreal master still per setup (camera height, identity at crop scale, geometry, facing, poses and wardrobe gated; ultrawide when the cast will not fit a portrait frame) → ONE multi-cut video clip derived from that master, QC'd on its cut boundaries → split into per-beat files once the creator passes it. Six sub-skills in order, four creator checkpoints. Invoke for a floor-plan-first scene or any scene with three or more people.
---

The blocking pipeline

A sanitized, abridged copy of the master skill and its six sub-skills, from the AI production pipeline on stolenduck.com. The full set is seven skill files of step lists with scripts and worked examples; this page keeps the method and every rule that came from a failure, and drops the paths, command lines and file conventions. Project, scene, character and colleague names are removed. It is current as of mid-October 2026 and was proven on three creator-passed scenes: a four-person exterior, a five-person interior and a twelve-person set.

What it is, and why

The keyframe pipeline on the main page generates one keyframe per shot and one clip per keyframe. It is right for a two-person conversation. With three or more people it fails on geography. The blocking pipeline sits beside it and reuses everything downstream. Three facts justify it.

1. Geography is what breaks. A video model given cast cards and a location plate puts the right faces in the wrong places: a character mirrored to the other pair, a party pushed across the set. A drawn floor plan with every camera on one sheet, plus ONE creator-approved photoreal master per setup, is the cheapest way found to pin positions. The clip derives its coverage from that master.
2. One multi-cut clip beats per-cut clips. One generation with timestamped hard cuts keeps light, wardrobe and geography identical across the cuts and lands every framing; single-cut clips copy the reference framing instead. Detected cuts land within about 0.25 to 0.7 seconds of the script, so splitting afterwards is safe.
3. Images are about ten times cheaper than clips. A 4K master costs 4 to 11 credits; a 12-second clip costs 84 at production resolution (36 at the test tier). Every check that can be made on a still (camera height, identity, geometry, facing, poses, wardrobe) is made on the master, and nothing goes to video until the creator has approved it.

An earlier version of this flow used generated line-art elevations per shot and independent generations per shot. It was rejected: the independent generations never agreed on geometry, and one generated plan came back rotated 90° against the location. What replaced it is below.

The phases

Checkpoints are one mechanism. Each is a delivered artifact plus the creator's decision in chat, written to one gate file, one line per decision: date, phase, item, files, the creator's words. The agent's own gate on a still is a pre-check: a fail means regenerate before the creator sees it. Automated clip QA only pre-flags. At every checkpoint the creator decides. Never record an approval the creator did not give, never fire a clip on a master without one, and never split a clip the creator has not passed. In testing, a clip was rendered on stills the agent had passed and the creator then rejected: credits spent for evidence only.

What each checkpoint delivers: (1) the shot list with its BLOCKING block, floor-plan text, cameras, exchanges and a look plan with outfit thumbnails; (2) the plate, the height crop, the cards and face panels, with any item still failing after regeneration as a pre-flag; (3) the master with its crop-scale face sheet and the floor plan beside it; (4) the clip with its QC sheet, a face-crop sheet from clip frames, the handheld metric, the audio spectrogram, the proposed split points and the pre-flags.

Hard rules

• The BLOCKING block is the single source of truth, pasted verbatim into every still and clip prompt, with people listed in frame order left to right.
• Camera text is projected from the floor plan, never freehand. A freehand "her shoulder at frame-left" contradicted the character's facing, and the model obeyed the text.
• One creator-approved master per setup; coverage is derived from it inside the clip. Per-cut stills are a repair tool only. Independent per-cut stills drift even when nobody moves.
• Every still and clip prompt carries IN FRAME and OUT OF FRAME lists (per cut, in a clip), a facing sentence per visible character, and the setup's GEOGRAPHY sentence, word for word the same in every prompt of the setup. A reworded geography sentence is a second, slightly different map.
• Neutral poses: in a still, everyone but the speaker; in a clip, listeners stay relaxed, hands open, with one small in-place reaction. The start frame's poses are mimicked for the whole clip.
• Identity is judged on a face crop beside the reference close-up, on every version, never on the full frame.
• Camera height is fixed by a composition statement plus a height reference scoped to height only. "Match exactly" is also matched for stillness and kills the handheld.
• House style repeated inside every cut; no music, no effects. To-camera interviews and first-person point-of-view scenes have their own camera language.
• No pre-line beats; the line starts on the cut. "…then pauses" only on the clip's final line.
• The duration follows from the lines, never the other way round (Phase 4).
• Every filled prompt is saved, per artifact and per version, append-only. A re-roll or a repair needs the exact text that produced the take.
• On a client timeout, rejoin the job; never resubmit blindly. The timeout only stops the client; the job is still running, and a resubmit pays twice if both finish.
• Regeneration cap. After three failed versions of the same still or clip on the same check, or four safety-filter rejections in a row, stop and bring the best version with its pre-flags to the creator. Past that point the prompt or a reference is the problem, and the creator decides what changes.
• Record every job's full id and every input upload id in one reference manifest; a truncated id makes a prompt unrecoverable.

Phase 1: the shot list

The BLOCKING block. One sentence per character in a fixed vocabulary: NAME frame-left/right/centre · foreground/midground/background · near <set piece> · facing frame-left/right/camera. Default to natural conversational blocking: partners shoulder to shoulder, two groups facing each other across a few metres. Depth in a portrait frame comes from the master camera looking diagonally along the group, not from staggering people; four staggered figures read as unnatural. Close with the axis, the camera side, and "Only written move: …" or "Nobody moves."

The floor-plan text and the GEOGRAPHY sentence. The same facts top-down, plus one sentence that names which side of any void each group stands on, where the key set piece is relative to the camera, and the spacing in metres. A group drifting across the void is the most common geometry error downstream, and without the spacing in words the image model bunches everyone along the nearest strong line in the plate.

The axis is the line through the parties, never the gap between them, and it is redrawn per exchange when three or more parties talk. Every camera used in one exchange sits on the same side of it. Drawing the axis down the gap put a camera over the line and flipped a reverse onto the far side. A small axis-check script reads the cameras and the exchanges and fails any camera that crosses the line or sits on or near it.

Singles. In a two-party exchange, every single is an over-the-shoulder from camera A (beside group 1, looking past their shoulders at group 2) or camera B (the reverse), so the partner is the foreground shoulder, spacing and screen direction never change, and every eyeline lands on someone. A third party standing between the groups, or a deeper reverse, is shot from a third camera or a named deep camera on the same side of the live axis; the third camera is also the clean wide for reactions. Never put a single's camera at the far end of the axis looking along it: tested, it spread the groups six metres apart, sent eyelines at nobody, and reversed depth between the two singles so the groups read as swapped. A speaker's camera sits on the side they face into, so it sees a frontal three-quarter, not a shoulder.

Camera height is eye level unless the script motivates another height. An unmotivated high angle on the opening wide breaks the handheld documentary register.

Every cut line carries shot size, camera letter, subject and facing, what is legitimately at the foreground edge, the expected background (what the plan puts along that sightline, never assumed), the in-frame and out-of-frame lists, business, and Only NAME speaks: "…". The first cut of every setup is an establishing shot that reveals each position.

Moves. A move that changes the positions others are blocked against (crossing to another group, an entrance or exit, sitting or standing) ends the setup: a new floor plan, a new master, and a seam clip between them. A move along a line that leaves everyone else in place, with both ends on the plan, stays in the clip and is named in the block. The twelve-person clip kept one character's walk down the centre line inside a single setup and passed.

Cast profiles, one block per character: basics, the look in this scene, and a voice with exactly one named accent. "Transatlantic" flipped between British and American from clip to clip.

Phase 2: references

Everything at 4K. The video model receives a 2048-pixel copy of every reference, so 1K cards and plates lose detail. A 4K card costs twice a 1K card, which is still cheaper than one clip re-roll.

The location plate is a 21:9 landscape, empty, rectilinear, telephoto-compressed, seen from an oblique three-quarter vantage so two faces of the architecture recede. Landscape, because it carries the side walls and depth the video model needs for reverse angles. Empty, because posed people in a plate get mimicked by the video. Most source photos of a location are head-on, and the plain prompt keeps a head-on photo head-on, so the plate prompt opens with an explicit camera move: "CAMERA POSITION IS THE WHOLE POINT OF THIS IMAGE. The reference shows this place HEAD-ON in one-point perspective. Do NOT reproduce that view. Physically move the camera to <a named corner>, standing eye height, and turn it to look DIAGONALLY across <the space> toward <a landmark>… Keep the SAME architecture, furniture and light as the reference; only the camera has moved." The plain prompt failed on the first location; the camera-move opening passed on two different image models.

The camera-height crop is a free 9:16 window cut from the plate at the master camera's vantage. It is passed to every master and clip as a height reference only: "match its camera HEIGHT and level lens; the camera itself is handheld."

Model cards: one per character per look, a 16:9 sheet in three equal vertical thirds: front view full body, the same person from behind full body at the same scale on the same ground line, and a large face close-up. One card replaces separate clothed-reference and back-of-head steps, and gives the video model a back view for over-the-shoulder cuts. Rules the cards taught:

• Pass the identity full-body and close-up, then the outfit image LAST, because the prompt says "wearing exactly the outfit shown in the last reference."
• The tail of the prompt stays whole, layout negatives included. Without "no panel borders or dividing lines," cards came back with dividers baked in and one with a wider face panel; the face-panel crop downstream depends on equal thirds. A thin divider between equal panels is harmless: pre-flag it, don't regenerate for it.
• A "cinematic, professional portrait photography" sentence from the template the cards started from is left out on purpose: it fights the house style, and no passing card used it.
• A character with no identity reference gets "a believable, specific face with natural asymmetry," plus named features, in place of the identity clause. The shared prompt fragments are never reassigned, or every card defined after the change silently loses its identity or outfit lock.
• Non-speaking characters can use an existing worn-outfit frame instead of a card. That saves the card job, not a reference slot.
• Safety-filter fallback chain for bare skin: cover-up wording or an opaque cover-up variant of the look, then the same prompt on a second image model at its maximum resolution, then a third model, then the creator. Rejections are free. One card was blocked twice on the default model and passed on the second.

Face panels are the right third of each card, cropped by the same definition the identity check uses, so the check judges exactly the panel the models receive. A face panel passed as its own reference fixed a drifted face that a face-only correction edit only half-fixed.

The reference manifest has one row per asset: the file, the full job id, and the baked-in facts (expression, accessories, hair, props, any text, the model and resolution if a fallback was used). Reuse an asset by its upload id to save re-uploads.

QA before the checkpoint, by opening every image: the plate oblique, verticals straight, horizon level, no people, no text; the height crop's horizon at standing eye height; the cards three equal panels of the same person in the right outfit, neutral pose, hands open.

Phase 3: the floor plan

The plan is drawn by code from a small JSON spec written from the floor-plan text, never generated by an image model. Its orientation is checked against the plate. Every camera of the setup is stamped on ONE sheet, each a black triangle with a dashed view cone aimed at what that camera frames and its letter. The axis check runs over the cameras and exchanges, and a camera that fails moves, the shot list's camera lines are amended, and the creator is told the approved shot list changed. The finished sheet gets a by-eye check, and is redrawn whenever a position or a camera changes. It is Reference 1 of every still and clip of the setup.

Phase 3a: the master still

One approved master per setup, never one still per cut. Eight rounds showed that independent per-cut stills never agree on geometry: the group's angle to the set's main feature, swapped partners and camera height all varied between generations, and with four people and nobody moving, three singles came back bunched, six metres apart, and swapped.

When the cast will not fit a 9:16 frame at eye level, make the master ultrawide. Forcing a large or spread cast into portrait makes the model lift the camera; three masters came back a touch high for exactly this reason. Generate a populated 21:9 master on the plate with the same references and gate, QA on the ultrawide where every face is large enough to check, crop the 9:16 start frame from it at the end the establishing shot starts on, and let the establishing shot traverse it. Validated on the twelve-person set: one image model took the whole cast with 16 references, and every cut of the 26-second clip held the geometry.

Reference order: the floor plan, the plate, the height crop, a card or worn-outfit frame per person in frame (left to right), then face panels for the speakers. Over a model's reference cap, drop the height crop first (the composition statement alone held the height on the twelve-person master), then keep face panels only for the main speakers.

The prompt is built from five fixed parts: a style block that ends on the GEOGRAPHY sentence; a composition statement; a reference map; the BLOCKING block; the master shot with its in and out lists and pose line; and a head count ("Exactly N people… no one added, no staff, no passers-by, no reflected figures. Nobody changes place; people only turn their heads."). Each version is appended as its own block and never edited, so the file reproduces every version.

The composition statement fixes camera height where adjectives did not: "a STANDING OPERATOR'S handheld camera at eye height with the LENS LEVEL, not tilted down: the horizon runs through the standing characters at eye height and sits in the upper third of the frame; the ceiling fills the top; the floor occupies only the lower quarter" (lower third for the ultrawide form).

Prompt rules, each the removal of a specific failure:

• List people left to right as they appear in the frame, never by depth. The image models lay a listed line out in reading order and ignore "nearest the camera"; two masters each reversed a line until both lines were written left to right.
• IN FRAME and OUT OF FRAME lists for people and set pieces, with where the out-of-frame ones are instead. Without them a fourth cast member and the plate's foreground table were imported into shots; with them both vanished first try.
• A facing sentence per visible character, naming who they look at and who they do not.
• Neutral poses on everyone not performing the scripted action: "arms loose, hands open, NOT clasped, NOT crossed, NOT on hips; only the speaker is posed mid-gesture." A master full of clasped hands yields a clip full of clasped hands.
• An identity line with the hard facts for anyone the model tends to drift: skin tone, hair or hat colour, a beard.
• The face panel as its own reference for every speaker.
• A character with a written walk stands at the walk's start in the still; the still is one moment.

The gate: five checks on every version, before the creator sees it.

1. Camera height. A standing operator, lens level, the horizon through the standing characters. A slightly elevated master is a fail, not a flag, because the clip starts on it and inherits the angle. Adjectives ("eye level", "not elevated") did not fix it across three passes; the composition statement plus the height crop did.
2. Identity at crop scale, every visible face, every version, on a sheet that pairs each card's close-up with the master's face. A master without its crop sheet is not reviewable. A face that passes at full-frame scale can be a different person at crop scale: one master turned a character into a man of a different ethnicity in the same shirt, a thumbnail check passed it, and the clip inherited it in every cut. The creator should never have to catch that. (A landscape reference photo cropped by the card rule comes out as background; crop such a photo to the face first.)
3. Geometry and spacing against the floor plan: order along each line, who is beside, behind or across from whom, which set piece each stands at, which side of the void, nothing imported, nobody missing, and the spacing matching the GEOGRAPHY sentence's metres.
4. Facing: each visible character looks at who they address; nobody at the lens or at nothing.
5. Poses and wardrobe: listeners neutral, only the speaker gestures, every outfit matches the look plan.

A fail is regenerated as the next version with the failing fact stated outright, in frame order. Identity drift alone is fixed by re-passing the face panel first.

Repair stills, only when a rendered clip's cut lands the wrong framing, geometry or face: a 9:16 still with the approved master as Reference 1 ("the same people in exactly the same places, same spacing, same light; ONLY the camera has moved to <letter>; reframe this scene, do not restage it"), gated and creator-approved like a master, then passed into the clip re-roll as one extra reference.

Phase 4: the video

Precondition: the master is creator-approved in the creator's own words, and its crop-scale identity sheet passed for every face. The start frame carries identity and geography into every cut; a drifted master makes a drifted clip regardless of cards, face panels or cast-lock text.

Pre-render audit: open the images. Before any video job, open every image the job attaches and list what is baked into each: expression, accessories, hair, props, any text. Diff that list against the prompt. Whatever is in the pixels wins at render time; a prompt line that contradicts an attached image is a block, not a lock. Fix it by changing the scene text or swapping the reference, never by adding a stronger negation. Also confirm every reference the prompt names is attached and vice versa, the order matches the numbering, and there is no lettering in any reference except the plan.

The start frame is a 9:16 crop of the approved master: for an ultrawide master, full height at the end the establishing shot starts on; for a portrait master, the master itself. The approved master as start frame tested better for multi-body geography than a reference-only render.

Duration budget. Per spoken line, the greater of 3.8 seconds or 0.4 seconds per word plus one second; a seven-word line with a physical action took about five seconds, so budget short performed lines toward that. Add about one second per party the establishing shot reveals (three seconds minimum) and 1.5 to 2 seconds per wordless reaction cut. Round the sum up to whole seconds, and set the clip length from it. Under-budgeted lines are not clipped: the model lets each line finish and pushes every later cut late, cumulatively; five lines in twelve seconds ran one to two seconds late by the end. Up to 26 seconds (the longest tested) is one clip; 27 to 30 is one clip with a "longer than tested" pre-flag or a split; anything longer splits into clips of 26 seconds or less at exchange boundaries. The creator-passed 26-second clip budgeted at 27 to 28 seconds and landed every detected cut within about 0.7 seconds of the script.

References for the clip: the floor plan, the approved master, the height crop, a card or worn-outfit frame for people in frame only, then face panels; the plate itself is not attached when the master is. A reference for someone off-frame makes the model insert them. Over the model's cap, keep in this priority: start frame, master, floor plan, speakers' cards and face panels, height crop, listeners' cards. Cloned voice references, attached as audio, stop accent drift.

The prompt, in this tested order:

1. Reference map. What each numbered reference is and governs: the plan as "the authoritative map of where everyone stands and which way they face," the master as "this exact moment… every cut is a view into THIS arrangement," the height crop as "camera HEIGHT and level lens only (the camera itself is handheld and alive)," then the cards in frame order and the face panels. Never a prompt token the tooling could read as a file path.
2. Scene state, then the BLOCKING block verbatim.
3. The cuts. Open with the bracketed count: "This clip has EXACTLY <n−1, in words> hard cuts and no others (<n> CUTs); each CUT is one continuous handheld shot with no cuts inside it." Then one paragraph per cut.
   • The establishing cut begins exactly on the start frame and never uses the word "pan": "The camera, staying at a distance from the characters, feels handheld throughout and energetically switches attention from <group 1> (<each one's action woven in>, for about one second) to <the next> …" The model deprioritises acting written after the camera block, so each character's action goes inline at the moment the camera reaches them.
   • Every later cut opens "HARD CUT. CUT k (t0–t1 s), camera <letter>," then the style tag, the shot size ("chest-up" or "waist-up", never "close-up", which creeps into a full-face push-in) with "hold this framing throughout the cut," the subject's facing, what is behind them per the plan, their neighbours, the in and out lists, and Only NAME speaks, starting on the cut: "…".
   • A PERFORMANCE line per on-camera character: an emotional verb phrase and one physical action that happen while the line is spoken, and a fixed lateral eyeline. A roving or absent gaze reads as "recompose" and drifts the subject.
   • Never a pre-line beat ("arriving in one breath", "a moment before"): it is staged literally as a gasp and a two-second pause before the first word.
   • A listener reaction only for someone already in frame: one small in-place reaction, a blink, a nod, a side-glance. A listener written off-frame cues a reveal: camera drift and random cuts.
   • Repeat the camera letter, the style tag and the acting direction inside every cut. The model treats each cut semi-independently and ignores a global intro; a cut that only inherited the global style came back locked-off or cinematic with shallow focus. The tag: "eye-level handheld, micro-shake, the operator's hands never perfectly still, clean digital, no bokeh, sharp across frame, natural light."
   • "…then pauses" only on the clip's final line; mid-clip it wedges dead air.
   • Where a character's facing changes during the clip, the cut's facing sentence wins over the plan's.
4. Camera rules: the GEOGRAPHY sentence again, word for word; every camera at standing eye level, never looking down at the floor, handheld in every cut; every camera on one side of each exchange's axis; eyelines never flip; nobody changes place between cuts, people only turn in place.
5. Head count and background life: "Exactly N people… no one added, no staff, no passers-by, no reflected figures. Non-speakers keep relaxed neutral poses, stay alive with small natural reactions and never look at the camera. No exaggerated faces."
6. The house style block, with a complete, stated list of where a multi-cut clip overrides the house style (no cuts inside a cut rather than none at all; "…then pauses" on the final line only; a written in-frame listener reaction; the tag repeated per cut). Do not add "slight drift", "follows action" or "fly-on-the-wall"; tested, they add excess camera motion.
7. Audio, text and effects negatives: diegetic sound only, and "ABSOLUTELY NO MUSIC of any kind: no score, no soundtrack, no underscore, no ambient pad, no drone, no stingers," because a one-line "no music" still produced a faint music bed. No transitions, flares, overlays, captions or logos.
8. Staging map only: the plan is a map; nothing drawn on it appears on screen. A single visible plan mark means the shot is wrong.
9. Start frame: begin exactly on it; the full master's positions, spacing, wardrobe and light hold for the whole clip.
10. Cast lock: each character's basics, this scene's look, and the speakers' voices with one named accent. Keep it short; a short lock holds identity better than over-description when the cards are attached.

In reference mode the start image is a strong reference, not a hard lock: it holds when prompt, plan and master agree and loses when they contradict each other. Consistency across all inputs is the mechanism.

Pronunciation: when a name or a coined word mispronounces, write the sound as ordinary separate words inside the spoken line, never a hyphenated respelling. Cheaper than a re-roll.

Throughput. A handful of concurrent video jobs is the plan's ceiling; count the unfinished ones before each submit. A 12-second clip takes three to five minutes, 26 seconds about ten. Never block a tier on one straggler: fire the rest and collect it later. Safety-filter rejections are free; the video filter judges bare skin in the output, so the levers are an opaque cover-up variant of a look and wider full-body framing. Before a re-roll, keep the old take under a take number.

Phase 4b: the QC sheet and the split

The QC sheet shows a frame just after and just before every scripted cut plus each cut's midpoint, labelled with the camera, the speaker and the line, with the cuts the detector found printed beside the scripted ones. One pass of continuity questions runs over it against the plan and the master.

Identity on clip frames is judged on the full-resolution midpoint of each cut, never on the sheet's thumbnails: a 300-pixel thumbnail of a 720p frame discards more than half its width, and the crop then judges an upscaled blur.

Handheld check. Background frame-to-frame motion on each static dialogue cut: a median under 2 is pre-flagged as locked-off, 2 to 9 is lively (tested lively range 3 to 8), above 9 is pre-flagged as too shaky. One clip measured 1.7 against another's 7.9 with identical style text in every cut. The difference was a height reference passed as "match its camera height and level lens exactly," which the model also matched for stillness. Scope that reference to height only and keep the live-camera cue in every cut.

Audio check. A spectrogram, looking for sustained horizontal harmonic bands under or between the lines of dialogue, then listening. Music present: regenerate with the audio block moved up beside the dialogue, or keep the picture and replace the bed in the edit.

Background direction check. On each cut, the background must match what the plan puts along that camera's sightline beyond the subject, judged against the shot list's expected-background line, never an assumption. A correct cut was once flagged because the prompt had said "house far behind" without projecting it.

Cut timing. Each scripted cut is matched, in one ordered pass, to a detected cut within 0.7 seconds; a scripted cut with no match is a pre-flag.

The split happens only after the creator passes the clip, at the detected cut times, not the scripted ones: a split at the scripted time carries frames of the neighbouring camera into the beat file. It is a frame-accurate re-encode, never a stream copy, because these clips have keyframes only at the cuts, and copy-splits came back audio-only or with pre-roll from the previous cut.

Costs

Stills: 4 credits at 4K on one image model, 11 at 4K high quality on the other. Video: about 7 credits per second at production resolution, 3 at the test tier. Rejected and failed jobs are not charged. A setup costs the plate, four credits per card, one line per master version, and the clip's seconds at the chosen tier. Worked examples: four people, three cuts, 12 seconds and two master versions is about 85 credits at the test tier and 133 at production; twelve people, 26 seconds, three cards plus worn-outfit frames and three master versions is about 134 and 238.