t3knobox/klik-anim-skill-creation/klik-remotion-method

Klik Motion Method

Prompt → discovery → visual storyboard → execution plan → build → QA. Keep it lean: this file is the spine; load ONE reference file for the stage you are in.

The gate model

Prose gates facts. Pixels gate aesthetics. A document can settle duration, codec, or an asset list; it cannot settle typography, composition or footage strategy — those are only real once seen. From the storyboard onward the user approves rendered frames, not descriptions, and klik-anim §A9 binds you first: never show the user anything you have not rendered and looked at yourself. A look-pass cannot catch durations — the arithmetic checks (reading holds, feasibility) run beside it, never instead of it.

The three-layer contract

• `klik-anim` — element craft laws (Part A, incl. §A8 only-the-user's-words,

§A9 look-before-showing, §A12 reading-hold arithmetic, §A14 device ledger, §A15 boundary continuity), Remotion layer (B), style hooks (C: pacing dial C0.5, style interview C2), pre-flight (Part D). Loaded for build/QA. Never overridden.

• `klik-cam` — camera movement: the camera dial C0, camera grammar

(signature variation, focus discipline, 3D hygiene), CAM checklist. Loaded whenever the piece's camera dial is above 1.

• This method — the orchestration: sequence, artifacts, gates, personas.

Never restate skill content in artifacts — cite section numbers.

Pipeline map

Reference files: references/discovery.md · references/storyboard.md · references/execution-plan.md · references/build-qa.md · references/reference-analysis.md (load whenever the user supplies a reference, any stage). Templates: templates/SSOT.template.md · templates/storyboard.template.html · templates/EXECUTION_PLAN.template.md · templates/PROGRESS.template.md. Tooling: scripts/analyze-audio.py (dependency-free: BPM/downbeat grid + the MUSIC MAP — sections, energy tiers, spectral lead, attack/sustain, per-source rhythm patterns, and JSON events/levels for deterministic reactive comps — consumed by klik-anim §A16 and klik-cam CAM-5. Optional stem mode via demucs, if installed, isolates kick/snare/hats/vocals/bass/ guitar/piano; degrades gracefully to 3-band analysis without it) · scripts/measure-motion.py (QA side of the same loop: frame-diffs a RENDER into its actual energy curve — event/rest contrast, floor-relative rest ratio vs klik-anim §A6, and per-bar correlation against the recorded arc/tier curve, klik-anim §A18 — run at the Stage-5 arithmetic pass).

The questioning sequence is law

Discovery follows the SSOT's numbered intake in order, always. The user may answer anything ahead of time (references early, script early) — record it in its section, then return to the sequence where they left off. The next question is always the SSOT's first ⬜ item. Ask in batches within a section; confirm anything inferred. Never generate an artifact the user already has — every section asks "do you have one?" before making one.

Context architecture (BMAD-style)

Artifacts are the context transfer — a handoff succeeds only if the document carries everything the next agent needs.

• Main conversation (Fable 5): discovery + storyboard — elicitation is

dialogue; personas frame the questioning, one conversation.

• Fresh context (Opus): execution-plan synthesis and build — inherit ONLY

the artifacts. The model/session switch IS the handoff.

• QA is always a separate context that did not author the code — an author

reviewing its own work rubber-stamps (role bleed). QA reads artifacts + render and judges.

Planning documents live on disk

• Everything is written to disk the moment it is drafted — an in-chat plan

cannot gate anything. Folder: user-assigned, else auto-create `klik/` at the project root (don't ask; say you did). Rendered gate material lives in klik/boards/ so a resumed session can see what was approved.

• `PROGRESS.md` (from template, created at discovery start) tracks pipeline

position, per-scene status, and the change log. Update at every gate, every scene built, every user "update progress". It is state; the SSOT is truth; the plan is contract — never merge them.

• Artifact headers carry Status: DRAFT | APPROVED. The user's approval flips

them — never flip one yourself. Post-approval changes are dated amendments.

Resuming

Read PROGRESS.md first, then verify it against the artifacts AND the code and renders — artifacts drift (a real resume found the plan saying AT-PROOF-GATE while PROGRESS said all-built). Reconcile every artifact to match reality before any new work. The ladder is authoritative:

1. No SSOT → Discovery, item 1.
2. SSOT with ⬜ intake items → Discovery, first ⬜.
3. SSOT approved, board missing or DRAFT → Storyboard.
4. Board approved, plan missing or DRAFT → Execution plan.
5. Plan approved, unbuilt scenes → Build (respect the proof gate).
6. All scenes built → QA.

House rules

• Cheap pixels early beat expensive pixels late. Stills at --scale=0.5

cost seconds; discovering the typography was wrong after a build costs the build. Never defer a visual decision to save a render — that economy has already failed once.

• Nothing is throwaway. Board mockups use the brand tokens; approved cards

become posed comps; the build animates those comps.

• One reference file in context at a time; during build, one scene row at

a time (plus klik-anim + SSOT).

• Iterate cheap: stills over clips, --scale=0.5, --frames=a-b;

full-res only at QA.

• Right-size. A 15s single-scene ask doesn't need the method — say so and

work under klik-anim (+ klik-cam) alone. Short videos compress the board, never skip it.

• Model split: planning phases on Fable 5; build on Opus — say so

at the plan gate, with the plan's token estimate (band, excludes iterations).

• Session limits: for ~30s+ videos the user WILL hit limits a few times —

warn at discovery start; artifacts + PROGRESS.md make resume lossless.