---
name: evidence-led-film
description: Plan, produce, revise, and review narration-led application films and product or founder stories using real work, physical motion, expressive sound, and incremental rendering. Use for Joe-style cinematic proof-of-work videos and their production retrospectives, not routine caption-only edits or IELTS episode formatting.
---

# Evidence-led film

Make the viewer want to take the intended next step through real work, clear visual hierarchy, and emotional momentum. This skill works in Claude Code and Codex with the project's existing video tools. It does not require a particular renderer, model, or plugin.

## Start from the current request

Determine whether the user wants planning, production, a bounded revision, or a retrospective. Planning-only requests remain planning-only. For an authorized revision, finish that scope without adding routine approval gates. Preserve explicit chapter locks and merge/publishing holds.

Locate the actual source project, accepted render, narration takes, asset references, and render commands before editing. A review HTML or flattened screenshot is not the source project. Inspect project instructions. Use existing context before asking questions; ask only for missing decisions that affect the work.

Read [design-and-motion.md](references/design-and-motion.md) for creative direction and [audio-and-rendering.md](references/audio-and-rendering.md) when editing or rendering. Use [production-sheet.md](assets/production-sheet.md) as a working template; fill relevant fields, not a bureaucratic checklist.

## Narrative before animation

1. State the intended audience, the judgment they should reach, and the next action. For an application, connect actual work to the role instead of building a generic personal montage. Do not invent employer endorsement or hiring guarantees.
2. Map narration into semantic beats. For each beat choose the real artifact, its role, the visual focus, and the change that occurs when the key phrase is spoken.
3. Separate approved facts/copy from suggestions. Use user-confirmed figures directly and consistently; do not restart authenticity audits. Missing material in the agent's records is not evidence that the user's work never happened.
4. Design entry, hold, and exit together. A beautiful still does not establish a good sequence. Avoid displaying all later proof before the narration reaches it.

## Joe's creative defaults

These are preferences learned from the Stripe film, subject to the current brief:
- Give important results and words real visual space. Tiny captions in an empty frame are not refinement.
- Let footage, products, correspondence, and real outputs carry meaning beyond spoken subtitles.
- Make motion connected: mass, acceleration, response, settling, and continuity. Breathing is subtle life, not perpetual jitter.
- Aim for optimistic, excited forward momentum. Sound should lift the story without obscuring the speaker.
- Select rich material with hierarchy. Repetition and six interchangeable cards do not create ambition.
- Preserve approved layouts when asked to animate them. Do not rewrite their copy or redesign them without need.
- Joe's established avatar treatment is circular, approximately 115% image fill; inspect the source geometry rather than blindly scaling every image.
- Keep an ambitious ending when the user chose one. Do not replace it with a smaller generic CTA.

Reuse these judgments, not Stripe colors, exact scenes, or OINP compositions in unrelated projects.

## Revision discipline

Before each edit, record **preserve / change / out of scope**, with the accepted version and exact affected range. The latest user correction overrides older suggestions. Do not average together incompatible historical versions.

Diagnose the failure category: narrative, layout, motion, sound, source media, or playback. Fix that category before adding polish. If a user likes a spatial entrance but asks for a new background, preserve the spatial entrance.

Keep shared content values, asset IDs, and timing in one source where the project supports it. Audit every affected occurrence including transition frames, embedded videos, cached imagery, and thumbnails. A fixed dashboard body can still flash an old number at a seam.

Render affected ranges plus transition handles first. Include adjacent chapters when their boundary changes. Reuse accepted rendered segments when technically compatible; do not rerender unrelated source films to replace two screenshots.

## Acceptance and delivery

Review normal-speed video with audio, then slow or frame-step suspicious seams. Use stills for composition and crop, not to certify motion or sound. If audiovisual inspection is unavailable, say what was measured and what remains unheard/unseen.

Check content legibility, motion continuity, full narration, inter-section sound consistency, musical energy, and media playback separately. Technical success is not creative approval.

Deliver the preview, exact changed scope, relevant verification, and remaining limitations. Preserve source files and approved versions. Publication, outreach, deployment, and merge remain separate actions.

## Retrospective and downstream reuse

After substantive feedback, record: observed failure, underlying cause, smallest successful correction, what must now stay fixed, and how to detect recurrence. Generalize the decision, not every numerical tuning value. Update reusable guidance only when asked.

For a companion site, reuse the accepted film, chapter times, selected stills, and original artifacts. Make evidence directly accessible without forcing a full rewatch. Do not automatically rebuild film scenes in WebGL or rerender a new montage.