Lofn Video & Animation — Codex-backed director pipeline

⚖️ AUTHORITY (2026-07-01): the .claude/skills/ twin of this skill is the CANONICAL policy source; this Codex mirror binds to it and to .agents/skills/lofn/EXECUTION.md §8 (Policy Deltas — golden-output quarantine, no-skip/NON-CANONICAL, itemized packet, per-pair variation angles, judge separation, the publish bar, gate mid-bands). On any disagreement, the .claude file wins.

Produces cinematic shot lists and animation prompts at Lofn competition grade. This one skill covers both modalities the user calls "video" and "animation". Depth lives in skills/video/ and skills/animator/; Codex is the engine (hybrid execution per .agents/skills/lofn/EXECUTION.md).

Rewritten 2026-08-15. This skill used to be a Veo 3.1 skill wearing a general name. Veo 3.1 is now the shortest frontier model on the menu (8s), native audio is table stakes everywhere, and prices span 12× per second. Model choice is now a Phase-0 decision, not a footnote. Landscape: output/analysis/2026-08-15_video-model-landscape.md. Rates and caps: vault/VIDEO_MODEL_DOSSIER.yaml.

Before you start

1. Confirm a Phase-0/1 packet exists (core_seed.md, 04_metaprompt.md, 05_pair_assignments.md, 06_director_handoff.md with the ICB / Panel Ledger, filled CREATIVE_CONTEXT.md). No packet → run lofn first.
2. PICK THE RENDERER NOW, BEFORE ANY PROMPT EXISTS (see Model-first below). Record it in the handoff. The prompt grammar depends on it.
3. Read the craft just-in-time:
   • skills/animator/SKILL.md — the four notation grammars, camera language, audio direction, the 6 motion archetypes (Pulse / Morph / Orbit / Parallax / Burst / Flow) and archetype chaining, loop logic, platform optimization.
   • vault/VIDEO_MODEL_DOSSIER.yaml — endpoints, caps, per-second rates, which models honour timed shot lists.
   • vault/VIDEO_RENDER_AUDIT.md — THE THREE-PASS BLIND RULE, mandatory on every render.
   • vault/DIRECTOR_QA_DEPTH_AUDIT.md — the Cinematic Somatic Gate + shot checklist.
4. Mode: Cinematic (multi-shot sequence / narrative clip) vs Animation (a loop/motion study). Same pipeline; Animation mode enforces loop logic (how frame 1 connects to the final frame).

---

⭐ MODEL-FIRST — the renderer is chosen at Phase 0, never after step 10

The house formula [CAMERA][SUBJECT][ACTION][SETTING][STYLE & AUDIO] is a Veo formula. It only ever looked universal because Veo was the only renderer. There are now four notation grammars and writing one model's prompt for another model wastes the run:

Grammar	Models	Write it as
cinematic	Veo 3.1 (all tiers), Gemini Omni Flash	Appearance and framing, camera front-loaded, one continuous shot per call
physical	Kling v3 standard/pro	Physical instructions to a camera operator and an actor on a real set — direction, level, duration, effort. Not a description of a painting.
declared-jobs	MiniMax H3 (≤7,000 chars)	Every reference gets a declared job; timed shot blocks; audio as a timeline of discrete events; explicit negations; identity locked by feature list
timed-ranges	Seedance 2.0 / 2.5	Timed ranges (0 to 3s … 3 to 7s); spoken dialogue in double quotes

Choosing: the dossier's role: field is the shortlist. Then —

• needs 4K native or a negative prompt → Veo 3.1 (accept the 8s cap)
• needs motion-heavy realism / physics → Kling v3
• needs character or style held across a set → MiniMax H3 or Seedance 2.x (reference-to-video)
• needs length → Seedance 2.5 (30s) or LTX-2.3 (20s)
• needs visuals cut to an existing track → LTX-2.3 audio-to-video ← the lofn-music bridge
• ⛔ there is no "drafting" choice — see TOP-LINE ONLY below. Divergence happens in text, for free.

⚠️ Every capability claim in the dossier is currently source: vendor-page — a claim, not an invoice. Anything model-specific in a run's docs is marked PROVISIONAL — VENDOR-DOCUMENTED, UNMEASURED until one of our own renders says otherwise.

---

⛔ TOP-LINE ONLY — and the ladder that follows from it

⭐ A BAD RENDER IS WORSE THAN NO RENDER (The Scientist, 2026-08-15). Cheap models are not a cheaper route to the same piece — they are a different, worse piece. Render only at the top line. The saving comes from getting the PROMPT right, which is what this ten-step, three-panel pipeline is for. It does not come from buying weaker inference.

Why this replaced the cheap-draft ladder: the three-pass blind rule detects incoherence, NOT averageness — and averageness is precisely how a cheap tier fails. A draft wave points money at the one failure mode we have no instrument for, and passes its own gate on the way to shipping the mean.

Top-line render targets: minimax-h3 · kling-v3-pro · veo-3.1 · seedance-2.5. Everything else stays in the dossier for costing and comparison and is not a render target for shipped work (preflight exit 5).

⛔ PROMPTS ARE NOT RENDERS

The 6 × 4 = 24 mandate is a DIVERGENCE requirement on PROMPTS, and prompts are free. It is not a render count and never was — the image rule has always read 24 prompts → top 12 renders. Conflating the two is what makes a pipeline unaffordable: 24 × H3 at 15s is $93.60, the whole account in one wave.

Diverge to 24 in text. Render the 1–3 that earned it.

The four rungs

Rung	What	Cost
1 · DIVERGE	24 prompts through steps 06–10	$0.00
2 · SELECT	rank, Somatic bloc, three-pass criteria applied to the text, The Scientist	$0.00
3 · TAKES	the 1–3 selected prompts × 2–4 takes each, at the top-line model	the only real spend
4 · POST	upscale / stitch / subtitle	pennies

⭐ TAKES, NOT DRAFTS. The pipeline removes prompt-caused regeneration. It cannot remove seed-caused regeneration — same prompt, same model, different take. One sample of a stochastic process is not a result, so budget takes of the right thing rather than drafts of everything.

Rules that carry over:

1. ⭐ RESOLUTION IS A FINISHING OPERATION, NEVER A GENERATION PARAMETER. Native 4K is ~$0.24/s; 1080p plus a post-upscale is ~$0.067/s — three and a half times cheaper for the same delivered resolution. And the event lives in the trajectories, not the pixels (point-light motion perception: observers read a walking human, and more, from ~a dozen moving dots). Buy duration and motion coherence. Upscale afterwards.
2. ⭐ A BAD WAVE IS AN ABORT, NOT A BUDGET REQUEST. If the takes come back wrong, the notation was wrong. Rewrite step 08 and re-select. Do not spend again to "see if it lands this time."

⚠️ The draft rung is now CONDITIONAL, and it is not the default. It is impossible for the model that matters — MiniMax H3 has a single tier ($0.26/s), so there is no same-family cheap rung, and a cross-family draft measures nothing (preflight exit 4). Permitted only where the same model has a genuinely cheaper tier and the notation is untested: veo-3.1-lite → veo-3.1 (8× spread) · seedance-2.5 @480p → @720p (2.1×). Everywhere else, the pipeline is the draft stage and it is free.

The cost gate (fails closed)

The default shape — one selected prompt, three takes, top-line model:

python3 scripts/video_cost_preflight.py --model minimax-h3 --duration 15 --count 1 --takes 3

--count is selected prompts, not the 24. Paste the output into the approval request.

Exit codes: 0 under ceiling · 2 OVER CEILING, refuse to dispatch · 3 bad input (unknown model, bad tier, duration outside the model's range — the estimate would be invalid) · 4 ladder families disagree without --bakeoff · 5 off-policy model — a render target outside policy.top_line_models, which needs --allow-non-top-line and a stated reason.

⛔ A PASS IS NOT PERMISSION. vault/AUTONOMY.md still owns the line: autonomous runs stop at drafts on disk; never render paid media, never publish, never spend without The Scientist. The preflight bounds an approved spend — it never authorises one. The deliverable of an autonomous run is paste-ready prompts + a costed render plan with the preflight output pasted in, and nothing else.

---

Execution (hybrid)

Coordinator 00–05 inline, then 6 pairs as parallel subagents for 06–10. At every agent start inject the full startup packet from EXECUTION.md §3 — CREATIVE_CONTEXT.md verbatim, the handoff/assignment files, the chosen renderer and its notation grammar, the current step contract, and the prior artifact where applicable. Default cardinality: 6 pairs × 4 = 24 shot-sets / loops → rank → top picks.

Coordinator steps (inline)

Step	File	Artifact
00	skills/video/steps/00_Generate_Video_Aesthetics_And_Genres.md	step00_aesthetics_and_genres.md
01	skills/video/steps/01_Generate_Video_Essence_And_Facets.md	step01_essence_and_facets.md
02	skills/video/steps/02_Generate_Video_Concepts.md	step02_concepts.md (12 concepts)
03	skills/video/steps/03_Generate_Video_Artist_And_Critique.md	step03_artist_and_critique.md
04	skills/video/steps/04_Generate_Video_Medium.md	step04_medium.md
05	skills/video/steps/05_Generate_Video_Refine_Medium.md	step05_refine_medium.md → 6 pairs

Per-pair steps (parallel subagents, one chain per pair)

Step	File	Per-pair artifact
06	skills/video/steps/06_Generate_Video_Facets.md	pair_{NN}_step06_facets.md
07	skills/video/steps/07_Generate_Video_Aspects_Traits.md	pair_{NN}_step07_aspects_traits.md (shot guide)
08	skills/video/steps/08_Generate_Video_Generation.md	pair_{NN}_step08_generation.md (4 shot-sets/loops)
09	skills/video/steps/09_Generate_Video_Artist_Refined.md	pair_{NN}_step09_artist_refined.md
10	skills/video/steps/10_Generate_Video_Revision_Synthesis.md	pair_{NN}_step10_revision_synthesis.md

---

The prompt contract (hard gate)

Every prompt names its renderer in the artifact frontmatter (renderer:, endpoint:, duration_s:, aspect:, tier:). A prompt with no named renderer is not a prompt, it is a paragraph.

Then write in that renderer's grammar. All grammars still owe the five concerns — camera, subject, action, setting, style & audio — they differ in how the concerns are expressed:

[CAMERA]  shot type + angle + movement       (cinematic grammar front-loads this)
[SUBJECT] specific, not generic              ("a woman in worn leather jacket, silver rings")
[ACTION]  exactly what happens
[SETTING] environment + time + weather/light
[STYLE & AUDIO]  aesthetic + explicit sound design

• Separate camera from action (each its own sentence). Describe negative space.
• Audio is directed explicitly — dialogue in quotes, SFX: prefix, Ambient: soundscape, Audio: music. On the declared-jobs and timed-ranges grammars, direct audio as a timeline of discrete events with named instrumentation, not a mood.
• Real film references allowed for look ("Terrence Malick golden hour"); no living-artist/actor likeness, no real victims. vault/HUMAN_SUBJECT_STANDARD.md binds.
• The 6 pairs use distinct camera grammar / archetypes — no two pairs default to the same move (the distinctiveness rule).

⭐ THE TAKE IS THE UNIT NOW, NOT THE CLIP

Seedance 2.x and MiniMax H3 cut internally from a timed shot list. The sequence has moved inside the generation call, which means editorial rhythm must be written down in advance:

[0–3s]  ECU: a hand closes around a moth.
[3–7s]  Cut wide: the same hand opens, empty, in a lit doorway.
[7–10s] The doorway light fails.
Audio: [0–3s] room tone, wing friction. [3s] one dry click. [7–10s] low drone, no music.

This is not lost editorial control — it is editorial control moved upstream, the way a long take is blocked before it is shot. But it only applies where the model honours it: check timed_shot_list: in the dossier. Veo 3.1 renders one continuous shot per call and will ignore your timings.

⭐ REFERENCE-TO-VIDEO IS HOW ONE SEAM GETS ENFORCED

L38 says a piece is made OF a material, never ABOUT materials: one dominant material (the world) + at most one intruder (the wound) + one seam, with the subject doing something across it. Reference-to-video makes that mechanical for the first time — assign every reference a declared job:

Use Image 1 for the world: the wet quarry, its light, its grain — this dominates every frame.
Use Image 2 for the intruder ONLY: the brass instrument. It is not in frame before 0:04.
The seam is at 0:04, where the hand carries it into the water.
Do not restyle Image 1 toward Image 2. No brass colour in the rock.

Same tool solves character/style consistency across a selected set. Caps: MiniMax H3 — 9 images, 3 videos, 3 audio (first 5 images free, then $0.08 each; reference video billed at $0.26/s of input). Seedance 2.5 — 50 files total (30/10/10), reference-to-video priced ×0.6.

---

QA & delivery

1. THE THREE-PASS BLIND RULE — vault/VIDEO_RENDER_AUDIT.md — on every render, before ranking. Pass A muted ("what happened?"), Pass B audio-only ("what happened?"), Pass C together and against the prompt. A and B name different events → FALSE WELD → repair or cut. Native audio and native picture are generated coincident, so a fused "does the audio serve the image" check cannot return NO — synchresis and the McGurk effect guarantee it. Writes <run-dir>/VIDEO_RENDER_AUDIT.md; no audit file → the render is UNAUDITED and may not be ranked, selected, or shipped.
2. Run lofn-qa with the Cinematic Somatic Gate (vault/DIRECTOR_QA_DEPTH_AUDIT.md).
3. Rank by: event legibility under mute → opening-second hook → motion clarity → audio-image cohesion (only for clips that passed A=B) → emotional legibility → loop integrity (animation).
4. THE DEDICATION IS A DELIVERABLE (vault/DEDICATION_STANDARD.md) — every shipped piece carries its ≤500-char dedication, measured not eyeballed. If the piece rides a named figure, credit the person who made it.
5. Save per skills/lofn-core/OUTPUT.md, frontmatter carrying renderer, endpoint, tier, duration, aspect, and the preflight cost line; INDEX last.
6. Do not call render tools without an approved, preflighted plan. Autonomous runs emit prompts + the costed plan and stop.