Moda Video Demo

moda-video-demo

PREREQUISITE — load moda-core once per session (step-0, write contract, free/metered map). Already loaded? Skip ahead. If you cannot load it, the non-negotiables: run moda doctor --json before anything; moda brand list before creating; writes that pin a revision use your last read's — on stale_revision, re-read and retry once (it heals); send the canvas link the moment it exists; stuck or failed? moda ask "<question>" — free and fast, never guess.

You are making a demo video of a feature by driving the app and recording it. There is no file to upload and nothing to screen-record by hand.

One command does the whole pipeline. Your job is to point it at the right target and to read what it tells you — not to rebuild it. Discovery, curation, validation, recording, pacing, the camera, captions, the critique loop and publishing all live in demo-capture.

The one rule that is not negotiable

Never drive a production target without explicit approval. The gate keys on the TARGET, never on the flow — do not try to judge whether a demo is "read-only". It cannot be judged reliably inside someone else's product, and a demo of clicking Delete is a legitimate demo.

A capture creates, renames, shares and deletes real things. Approval comes before execution, not after.

1 · Find the app — do not ask the human yet

Work out how to reach it from the repo itself, stopping at the first rung that works:

1. Already running? Probe the usual ports (curl -s -o /dev/null -w '%{http_code}').
2. A start command? package.json scripts, Makefile targets, Procfile, README quickstart.
3. How do its own tests sign in? *.setup.ts, playwright.config.*, an e2e/ directory. A programmatic sign-in the test suite already uses is reproducible and safe.
4. Only now, surface what you could not establish and ask.

Prefer the app's own test auth over the developer's live browser session: more reproducible, dramatically safer, and the pass that finds the login usually finds the seed command, which is where determinism comes from.

Report the target, the auth strategy, and anything the repo did NOT establish.

2 · Find the tool, then check the machine

It installs with this skill, as demo-capture/ beside this file. Locate it once and keep the path:

DC=
for c in "$CLAUDE_PLUGIN_ROOT/skills/moda-video-demo/demo-capture" \
         ~/.claude/skills/moda-video-demo/demo-capture \
         ~/.agents/skills/moda-video-demo/demo-capture \
         ~/.cursor/skills/moda-video-demo/demo-capture \
         .claude/skills/moda-video-demo/demo-capture \
         ~/.claude/plugins/*/moda*/skills/moda-video-demo/demo-capture \
         ~/.claude/plugins/*/*/moda*/*/skills/moda-video-demo/demo-capture \
         ./moda-cli/skills/moda-video-demo/demo-capture; do
  [ -d "$c" ] && DC=$c && break
done
if [ -z "$DC" ]; then
  echo "demo-capture not found — it installs beside this SKILL.md; reinstall the skill"
else
  echo "DC=$DC"
  node "$DC/doctor.mjs"
fi

Names what is missing — node, ffmpeg, the browser, both CLIs, whether moda is signed in — before anything is recorded. Three states per row: present, missing, or unknown. Fix what it names; do not start a capture around a gap.

3 · Run it

node <DC>/run.mjs "<what to demonstrate>" <url> --name <slug> [--no-auth] \
  [--attempts 2] [--publish "<Title>"]

<DC> is the literal path step 2 printed — paste it in. Each Bash call is a fresh shell, so $DC from step 2 does not survive into this one.

--attempts 2 lets it re-discover when the first flow is not worth filming. Without --publish it stops after the critique and prints the command to publish, which is the right shape when you want to look first.

Read its report rather than summarising it. Everything below is measured, and each line is either a number or an explicit "not measured" — never a silent pass:

A run that refuses to record is doing its job. Do not work around it by hand-writing a flow — find a better goal, or say the app has no good demo for this one. That is a real answer.

4 · Say what it actually produced

Give the human the canvas URL and the mp4 path, the score, and the findings that survived. Do not oversell a mediocre take. If the report says two thirds of the video is a still image, say that. The score is a proxy for polish, not for whether the flow was worth filming — a clean video of a pointless sequence can outscore a rough video of a good one, so trust the flow findings over the number.

Known limits — say these plainly, do not oversell

• A long wait is compressed because dead air is boring, not to fit a cap. Idle gaps play at 6×, and each pacing fix adds 3 up to 14×, keeping a beat at each end and never touching the final reveal — measured, a 109s prompt-to-design run became a 39s video. Length is rarely the constraint: mp4 exports at 18000 frames / 600 s, i.e. ten minutes at 30fps, and rejects rather than truncating past it.
• Punch-ins near a frame edge do not centre — the camera stops flush rather than exposing the page behind the recording. Reported as zoom_framing_clamped.
• No brand kit is applied yet.
• A studio checkout only makes the camera round trip local. Without one the loop plans the camera on the server, through the same planner publish uses — free, and writing nothing — so the zoom checks are measured and the camera is tuned either way. The checkout saves a network call per round, nothing more.

Not this skill

Footage the user already has → there is no upload lane; say so. Animating a logo or a design → moda-video-clip. Motion on an existing canvas → moda-video-motion.