agentleFS
ryanportfolioryanportfolio/Harness-Firmware/.agents/skills/wow-animation/SKILL.md

wow-animation

Skill1 starsChanged 9 days agoMITView on GitHub

$wow-animation <subject> [notes], or start a new shareable promo, explainer or trailer animation. Kickoff + orchestration; craft → motion-design.

  • Commits and pushes

5 other projects use a version of this, 4 of them edited.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

4 rules. 1 is common, 0 are distinctive, and 3 are very specific to this project.

  • 1

    Common

    Most of the 4 edited versions we read keep these.

  • 0

    Distinctive

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Check which agent tools this session exposes first.

    line 44
  2. Very specific:

    Ship: Commit, push, PR (gh pr create --head <branch>), the Opus review per AGENTS.md, fix surviving findings.

    line 62
  3. Very specific:

    Every on-screen string and number comes from facts.md.

    line 72
  4. Common:

    No audio of any kind unless the user's notes asked for sound.

    line 74

“Common” is measured: kept by at least 6 in 10 of the 4 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Wow animation

Kickoff brief and orchestration for one shareable JavaScript animation about a real subject. This skill decides what to make, grounds it in the real thing, runs the agents and the review gates. The motion-design skill owns story craft, timing rules, the frame(t) engine, verification and export. Never restate or override its rules here; load it and follow it.

Argument

$wow-animation <subject> [notes]. Subject is a name, URL, repo path or any mix (example.com, <game name> <store page URL>). Everything after the subject is notes: extra goals, must-show items, and overrides of the defaults below. No subject → ask once, then go.

The brief (the job every run carries)

Build a shareable JavaScript animation that:

  1. Makes a stranger understand what the subject is and why to use it or play it.
  2. Is worth watching as information on its own, even for someone who never clicks through.
  3. Matches the subject's existing look without being limited by it. Creativity first: if it wows you, it will probably land with others.
  4. Is gorgeous enough that people want to watch it to the end.

Ground every frame in the real thing: real assets restyled, real product strings, real numbers. Full text for agent briefs: brief.md.

Defaults (notes override any of these)

DefaultOverride examples
16:9 desktop and shared-video viewing; no portrait or phone cut"also portrait", "phone first"
Silent. No music, no effects, no AudioContext, no audio stream in the export; captions carry the story. This beats motion-design's music-led trailer cutting: cut on visual beats instead. Never add sound because it seems to fit"sound", "music", "audio" → follow score.md
Deterministic frame(t) with seek/renderAt and flash limits (per motion-design). No reduced-motion version"reduced motion" → build one per motion-design
No periods in headings or large display text (statements, taglines, end slate); product UI copy keeps its own punctuationnone expected
No MP4 or video export until every wow-loop round passes; export is the last step"export a draft now"
One approval checkpoint: beat sheet plus animatic, then run to the end"autonomous" skips it; "check in after each scene" adds more
Maximum parallel fan-out (see Parallelize)"no subagents", "keep it small"
$opus-fullreview after each major step; the AGENTS.md Opus review before any merge"skip review" (the before-merge review stays: AGENTS.md)
Games: trailer structure per motion-design (gameplay fast, hook early, logo on end slate), real enemies, bosses and systems"teaser only", "no bosses"
Commit and push on a feature branch, open a PR when done"local only"

Parallelize

Speed comes from fan-out.

  • Check which agent tools this session exposes first; multi_agent = true alone does not prove they exist. No agent tools → say so, run steps in the main thread one after another, and never claim fresh-context critics or directors ran.
  • Spawn every independent agent of a step concurrently, then wait for the results each next step needs. Before each dispatch, ask what else could start now.
  • Subject research agents: what the subject is, what it does, how it looks, its copy, assets, store or marketing pages, code facts. Split by source (live site, repo copy, design tokens and assets, feature code, store page, sprites, enemy and boss code) so each agent owns one slice with no overlap. Six or more recon agents is normal for a large subject. A lighter model is fine here when the tools let you pick one.
  • Judgment calls (animator research, direction pitches, hero scenes, restyled assets, critics) run on the strongest available model. Honor explicit user model choices.
  • Steps 2 and 3 start together. Director pitches run in parallel. Build modules with disjoint files run in parallel. Critics of one wow-loop round run in parallel.
  • Serialize only what must be: one writer per file set, one owner per browser profile, performance measurements, and anything that needs a prior result.
  • The orchestrator stays free: it merges reports, keeps facts.md and state current, and dispatches the next wave; it does not do recon itself.

Workflow

  1. Locate and branch. Find the subject's repo: current repo, then sibling folders of the current checkout, then gh repo list. Verify on GitHub, never the local checkout; branch a fresh worktree off fetched origin/<default>. Put the animation where the repo's conventions say; no obvious home → ask once. No repo access → public sources only, say so, mark every claim unverified. Sandbox: browser launches, the local server and fetching public pages need process and network access. Request approval for those commands; never loosen the sandbox globally or change Codex configuration to get it.
  2. Ground (fan out). Parallel recon agents, one per source slice, spawned alongside step 3, build facts.md next to the animation: every on-screen string, number, feature, enemy and asset with its source (file:line or URL). Checklists per subject type: grounding.md. A claim with no source gets cut, never guessed.
  3. Research the masters (runs with step 2). One agent searches for the most respected animators and motion designers as of today's date and what they teach. It writes inspiration.md with sources, flags what is new against motion-design's references/principles.md, and names 2 to 3 teachings this piece will apply. Propose additions to the motion-design skill in the handover; never edit it mid-run.
  4. Direct. Load $motion-design. Two or three directors each pitch a visual direction plus a one-sentence feeling-and-job line, the piece's type (per motion-design's references/critique.md) and a timestamped beat sheet, built from facts.md and inspiration.md. Pick the strongest (or graft), build the animatic, show the user beat sheet plus animatic, wait for OK unless notes say "autonomous". Write the chosen brief, direction, feeling-and-job line, Type: line and beat sheet to BRIEF.md beside the source and commit it.
  5. Build. Scene modules with disjoint files, one writer each; strongest model on hero scenes and restyled assets. $opus-fullreview on the build.
  6. Wow loop. Run $wow-loop (absent → say so and run motion-design's verification by hand). Seed its contract from the four brief goals plus the defaults above as named checks, including one "stranger test" check: after one natural playback a critic states what the subject is and why to use it, and it matches facts.md. Add the critique questions for the BRIEF.md type as checks, with the scripted reading-time check and the protected strongest moment, per motion-design's references/critique.md. The stranger test doubles as its blind q-point: that critic gets no kickoff brief, BRIEF.md or contract text for the check, and its sentence must match both facts.md and the feeling-and-job line. Critics judge the live JS animation at natural speed, plus stills and contact sheets. Rounds continue until the contract passes or the budget runs out. $opus-fullreview after the loop.
  7. Export last. Video per motion-design's references/verify-export.md, only after the loop passes. Silent run: confirm with ffprobe that the file has no audio stream before calling the export done.
  8. Ship. Commit, push, PR (gh pr create --head <branch>), the Opus review per AGENTS.md, fix surviving findings.

$opus-fullreview needs an authenticated Claude CLI; unavailable → run $impartial-review and say which ran (it needs fresh independent agents or an authenticated Codex CLI; never substitute a self-review). Claude CLI unavailable for the before-merge review → say so, stop before merge.

Rules every subagent brief carries

Subagents never see this skill. Paste these into each brief that renders, captures or writes copy:

  • Headed Chrome on the real GPU, offscreen and muted: launchPlacedChrome() if the repo has it, with CHROME_PLACE=offscreen set in the command (Codex does not inherit Claude Code's environment), else --window-position=-2400,-2400 --mute-audio on the first window. Never headless for motion claims, never minimized.
  • Screenshots, stills and contact sheets go to the machine's screenshot folder when the user's instructions name one, keeping the relative path they would have had in .tmp/; otherwise .tmp/. Pass the exact path into the brief and report where they went.
  • Every on-screen string and number comes from facts.md. No periods in display text. On-screen copy passes the writing skill (no em dashes, no puffery).
  • Flash limits per motion-design. Reduced motion only when the owner asks.
  • No audio of any kind unless the user's notes asked for sound. When they did, paste the rules from score.md into every brief that writes or reviews audio.

Pitfalls

  • Starting pixels before facts.md exists: the piece drifts to generic stock visuals and invented features.
  • Faithful copies of sprites or UI that explain nothing to a stranger. Restyle the real asset, then make its rule readable.
  • Exporting MP4 to show progress: critics then judge compression, and each fix costs a re-export.
  • Critics judging only stills. Motion verdicts need one full natural playback.
  • Duplicating motion-design rules into briefs from memory instead of loading the skill.
  • Adding music because a trailer "needs" it. Sound only when the user asked.
  • Calling audio good from code review or a subagent's say-so. No agent can hear it; the checks in score.md plus the user's listen decide.

Handover

PR link, where the animation lives and how to open it, screenshot folder path, wow-loop outcome per check, each review run (which skill, findings that survived verification, what was fixed or waived), proposed additions to motion-design from inspiration.md, open items.

Examples

$wow-animation example.com
$wow-animation <game name> <store page URL> show the final bosses and the upgrade systems

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.