/open-prompt-showreel

name: open-prompt-showreel
description: Make a design-led motion-graphics video (showreel, eval comparison, feature or bug-hunt story) from one open prompt. Claude Opus at max effort designs, codes and renders the whole thing itself in a scratch dir, with no template and no repo PR. Use for any "make a video for X" / "do a showreel for Y" / "a little comparison video" ask aimed at X, LinkedIn or a launch post. Covers the Lead brief template, grounding every on-screen number in a real source, the real Agent Swarm end logo, CC0/CC BY music with credit, render specs (16:9 + 4:5), the contact-sheet self-critique loop, agent-fs delivery, feedback rounds and crash-safe retries. Not for fixed-template creatives or for replicating a reference clip.
metadata:
  type: procedure

Open-prompt showreel videos

The approach: give a strong model on max effort one open creative prompt plus hard content rules, and let it design, code and render the whole video itself. Template pipelines produce generic cuts. This one produced the showreels and the TLA+ races cuts that the team liked.

When to use which skill

| Ask | Skill | |---|---| | A video that should look designed: showreel, model/eval comparison, "we found N bugs", launch hype | **this skill** | | A standard on-brand card or clip from a fixed template (release, proof stat, quote, OG image) | a template toolkit, if your swarm has one | | Rebuild the feel of a reference clip someone linked | `motion-design-video-analysis`, then `motion-design-replication` | | Repo-committed Remotion video (README hero, landing demo, case-study clip) | `video-generation` |

If the ask says "like the showreel", "your own design" or "go all out", or someone complained a cut looks generic, it is this skill.

Model, effort, harness

• The maker runs **model `claude-opus-5-5`, effort `max`**, on a **Claude Code harness worker** (`provider: claude`). Set both on the task.
• Other harnesses were not tried for this workflow. Do not route to them without a test run.
• Every brief ends with a gate: "If you are not on claude-opus-5-5, stop and say so instead of rendering." The maker checks this first.
• The worker needs Playwright + Chromium, ffmpeg, python3 with numpy, and write access to agent-fs. The community template's `setup.sh` checks and installs all of them (numpy goes in a venv; the stock image does not ship it).

Lead brief template

One task, one open prompt. Do not write a storyboard, shot list, palette or template for the maker; that is what made earlier cuts generic.

GOAL: <one line: what the video is and where it goes (X / LinkedIn / internal)>.
Made by you on claude-opus-5-5, max effort, ad hoc, your own design.

<Requester>, verbatim: "<the ask>"

Your prompt: make a dynamic <15 | 15-30> second motion graphics video <about SUBJECT>.
Make it your showreel: show what an incredible motion designer you are. Go all out.

Content rule: every number or claim on screen comes from <SOURCE: eval writeup path /
test file on origin/main / PR numbers> and matches it exactly. Anything not in the
source stays off screen. Fewer numbers is fine: design-led, not a table.

How:
- Ad hoc: scratch dir, any tools or code you like. No repo, no PR, no shared template.
- Real brand: Agent Swarm logo = origin/main apps/ui/public/logo.png (verify vs agent-swarm.dev/logo.png). Never draw your own mark.
- Look at a contact sheet at least 3 times, fixing what looks weak each round; in the last round check every on-screen number against the source and look at the end-card frame.
- Music optional; CC0 or CC BY only (no NC, no ND, no SA); short credit on the end card; in your output the full attribution from music/current.json (source link, license link, changes made).
- If you are not on claude-opus-5-5, stop and say so.
- Nothing is owed to a human by you; Lead reviews and relays.

Deliver to agent-fs thoughts/<maker-agent-id>/videos/<slug>/: <1920x1080 [+ 1080x1350]> h264 yuv420p mp4,
contact sheet(s), facts.json.
Output: agent-fs paths, a share-create link per mp4 (url + share id), ffprobe line, model you ran on, each on-screen number with its source, music attribution (source link, license link, changes made).

Optional style add-on for a calmer, premium cut (the "Apple framework"): 3 colors, one background, one font, no HUD/grain; one shot = one idea; ~3 s open and ~2 s held end; key object centered; eased, overlapping keyframes; no hard cuts; ~108 BPM, phrase hit on the first product frame; no SFX. Add it only when asked.

Maker procedure

1. Scratch dir first, and say where it is

`mkdir /workspace/personal/<slug>` and store-progress with the path in the first minute. `/workspace/personal` survives a container restart; the session does not. A retry that found the scratch dir of its crashed predecessor delivered in 8 minutes.

2. Facts before design

Write a small script (`facts.ts`, run with `bun facts.ts > facts.json`) that reads every on-screen fact live from its source right before render, and have the reel read `facts.json` instead of hard-coding numbers.

• Code facts: `git fetch origin main`, read the file at `origin/main` (for example count `test.failing` vs `test` in `src/tests/workflow-tla-races.test.ts`), `gh pr view` / `gh pr list` for PR states. Record the short SHA in facts.json.
• Eval facts: the writeup the eval task produced on agent-fs, and the run id. Copy numbers exactly.
• A fact turns "green" only if main says so. Unmerged stays in its old state, even if the brief expects otherwise.
• If the source disagrees with the brief, the source wins; say so in the output (the TLA+ v3 brief listed four merged PRs; main also had a fifth, so the reel said "ALL 10 FIXED ON MAIN").
• Drop any claim the source does not support.
• If the source is not ready yet (eval still running, PRs still merging), build the design on draft facts, then `defer-task` with `wakeOn:{event:"settled",taskIds:[...]}` and re-read the facts on wake. Never ship draf