# Capture > Autonomously produce browser-extension launch assets with take-a-repo, then present the technically verified final media for explicit user approval. Use for CWS/YouTube promo video, X video, YouTube Shorts, store screenshots, listing/privacy evidence, or channel variants. Infer mechanical channel settings, capture, validate, fix, and retry without interrupting the user; bind Approve or Request changes to the exact final file digest. - Skill: `heznpc/capture` (Agent Skill) - Install (CLI): `npx skillmds@latest add heznpc/capture` - Raw SKILL.md: https://api.skillmd.com/api/skills/heznpc/capture/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: heznpc (https://skillmd.com/u/heznpc) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/heznpc/capture --- # Produce and approve launch assets with take-a-repo This skill is the browser-specific path. When completing a product with native, CLI or API evidence, use the bundled `launch-proof` skill and `docs/evidence.md`; do not force the product into a web page. take-a-repo drives the repo's **built** extension with Playwright and writes assets into the config's `outDir` (default `store-assets/`). A successful run doubles as a real-bundle smoke test — a screenshot or clip only appears if that feature rendered from the shipped code. By default, it also writes a handoff pack: `storyboard.json`, `captions.json`, and `take-a-repo-manifest.json`. ## Autonomous workflow 1. **Translate intent into targets** — infer channel profiles from the user's campaign request. Supported targets are `cws-youtube`, `x`, and `youtube-shorts`. Do not ask the user to choose viewport, codec, duration, thumbnail timing, or editor. 2. **Preconditions** — the repo has a `take-a-repo.config.js`; Chromium is installed (`npm exec -- playwright install chromium`, one-time); the config's `build` command succeeds. 3. **Create or update one story** — keep product actions and captions in one demo and declare channel variants through `targets`: ```js demo: { name: 'skillbridge', targets: ['cws-youtube', 'x'], captions: [ { at: 0.5, text: 'Translate the lesson in place' }, { at: 18, text: 'Restore the original anytime' }, ], async run({ page, env, demo, target }) { /* one reusable story */ }, } ``` take-a-repo expands target-specific names, viewport, H.264 MP4, 30-second cap, poster frame, and caption treatment. The `youtube-shorts` profile uses three-word outline focus chunks with an animated current-word highlight and a visual-guide-safe left/bottom placement. Use `targetOptions.` only when the shared story genuinely needs target-specific framing or caption tuning. If a desktop UI does not reflow at 720×1280, give Shorts a focused `run` override and fixture layout that removes secondary panels and enlarges the action/result. Never squeeze the complete desktop story into the vertical frame. For every localized campaign variant, declare `captionOptions.typography` with the exact locale and project-local licensed font files. Prefer a font family designed for that script; include fallback faces until glyph QA is complete. Do not rely on the operator machine's system fonts for a publishable multilingual asset. 4. **Run attempt 1** (from the repo, or pass its path): ```bash take-a-repo --json --attempt 1 take-a-repo --json --attempt 1 ``` If repeated composition fixes remain unresolved and the repo declares `config.calibration`, start `take-a-repo --calibrate`. Keep adjustments inside its authored presets, bounded framing/caption controls, and three protected regions. Save the profile, trigger the real recapture, and continue only from its resulting `publish-ready` or structured `needs-fix` state. Do not ask the user to diagnose composition or operate the controls. Once technical QA passes, open `take-a-repo --campaign` for the user's final media decision. Before npm publication, run the source checkout with `node bin/take-a-repo.js --json --attempt 1`, or use a project wrapper such as `npm run capture:store -- --json`. Useful flags: `--scene ` (one story, expanded variant, static scene, `description`, or `privacy`), `--target ` (one channel target), `--no-video` (skip the screencast), `--mp4` (also emit an H.264 mp4 of the demo — needs ffmpeg on PATH or `TAKE_A_REPO_FFMPEG`), `--no-build` (reuse an existing build). 5. **Read the result** — stdout is exactly one JSON object: `{ "ok": true, "status": "awaiting-approval", "machineStatus": "publish-ready", "outDir": "...", "manifest": "/abs/path/take-a-repo-manifest.json", "produced": [...] }`. Progress logs go to stderr in `--json` mode. Read `handoff.automation` for technical repair work and `handoff.approval` for the user decision. Do not use the legacy compatibility review summary. 6. **Fix, review, and publish through the gate**: - `needs-fix`: apply every `automation.actions[]` item whose owner is `agent`, edit the config, then rerun `automation.retryScenes[]` with `--attempt 2`. Repeat through `automation.maxAttempts`. - `blocked`: automated attempts are exhausted. Report only the concrete technical blocker and attempted fixes; ask for technical input. - Before presenting any video, inspect the final composite against its audience, objective and beat intent: evidence, composition, legibility, pacing and continuity. Caption pixels and codec checks cannot judge those. In evidence/production configs, author `editorial` and run `production review-context`, then `production review --report ` with inspected frame references and timestamped findings. `editorial-review-required` means agent work remains; another identical render does not resolve it. See `docs/evidence.md` for the report contract and detail sampling. - `awaiting-approval`: technical QA and agent critique passed. Open the Campaign Dashboard and present the rendered candidate to the user. Keep the Calibrator under Advanced for agent-owned composition work. Do not approve on the user's behalf. - `changes-requested`: read the digest-bound decision note, implement it as the next agent-owned edit, recapture, and return the new candidate for another decision. - `approved`: the exact recorded digest passed user review. An authorized uploader may publish that digest; any recapture or profile edit invalidates the decision. - `not-requested`: legacy capture mode; no channel target was configured. 7. **On runtime failure** — exit code `2` = usage/no config found, `1` = runtime failure; stdout still carries the single JSON payload `{ "ok": false, "error": … }`. Common causes: build failure, Chromium not installed, an unknown `--scene`, or a scene's wait timing out (feature didn't render). ## Notes - Runs the full-Chromium channel; headless works (`TAKE_A_REPO_HEADED=0 take-a-repo …` — verified, video included) and is the mode to use in CI. Headed-under-xvfb is unreliable on CI runners — don't use it. - Scenes are the repo's own config — to change *what* is captured, edit `take-a-repo.config.js`, not take-a-repo. - `description.from` may point to `STORE_LISTING.md` for copy only or to `product.manifest.json` for shared listing + privacy disclosure inputs. `privacy-disclosure.md` is a worksheet for store review, not legal policy text. - CWS assets and SNS demo clips have different jobs. For X/SNS clips, prefer `demo.preset: 'sns-video'` (`1280×720`), `demo.mp4: { crf: 18 }`, 20-40 seconds, short captions, first-result-within-3-seconds, and a story that demonstrates the requested behavior. These are recommendations; do not pad short cuts or force safety/restore copy without a visible reason. Use `sns-twitter` (`1200×675`) for static X card images. - Demo configs can use timed `demo.captions` plus the helper passed to `demo.run`: `demo.caption(text)`, `demo.step(text, async () => { ... })`, `demo.wait(ms)`, `demo.click(selectorOrLocator, { moveMs, beforeMs, holdMs })`, and `demo.select(selectorOrLocator, value, { openMs, holdMs })`. Captions, arrow-pointer clicks, and mirrored native-select options render as DOM overlays during recording and avoid the top-left disclaimer badge. Always use `demo.select()` for a native `