A collection of agent skills — self-contained capability packages that Claude Code, Cursor, and other skill-aware agents load on demand.
Each top-level directory is one skill. A skill is a SKILL.md describing when and how to use it,
plus whatever code, examples, and dependencies it needs to actually do the work.
agent-skills/
├── recording-web-walkthroughs/ # narrated MP4 walkthroughs of web apps
│ ├── SKILL.md # frontmatter (name, description) + the instructions
│ ├── run.mjs # CLI entry point
│ ├── lib/engine.mjs # narrate → record → compose
│ ├── example/ # runnable example, no network needed
│ └── package.json # dependencies scoped to this skill
└── publishing-android-to-play/ # ship an Android app to Play internal testing
├── SKILL.md # the detect → gate → step-loop protocol
├── verify.sh # release gates, PASS/FAIL per line
├── assets/ # Expo config plugin for release signing
└── references/ # per-framework build paths, Console steps
SKILL.md frontmatter is the contract with the agent: name must match the directory, and
description is what the agent matches a user request against, so it lists the trigger phrasings
explicitly rather than summarising.
Dependencies are per-skill, never hoisted to the root. Each skill installs and runs from its own directory so it can be copied out on its own and still work.
Turns a scripted browser session into a narrated MP4 with a visible cursor, smooth scrolling, and captions. Playwright records, the OS text-to-speech narrates, ffmpeg composes — entirely on-device, no cloud services.
The design centres on one idea: each step is a chunk owning one narration clip and one video segment, both built to the same duration. Sync holds by construction, with no global offsets to drift. Audio is never retimed — only the picture stretches to meet the voice.
cd recording-web-walkthroughs
npm install
npx playwright install chromium # one-time, ~100MB
node run.mjs example/demo.storyboard.mjs --out example/demo.mp4
node run.mjs /abs/my.storyboard.mjs --keep --headed # watch it, keep intermediatesSee recording-web-walkthroughs/SKILL.md for the storyboard
format, the ui helpers, within-step timing, and a table of non-obvious failures.
Requirements and constraints
- macOS by default. Narration uses the built-in
saycommand. Other platforms need attsCommandin the storyboard config (Piper and Kokoro both run offline). ffmpegandffprobeonPATH, used for every compose and verification step.- Node 20.11+ for ESM and
import.meta.dirname; developed against Node 22. - Run from the skill directory —
node_moduleslives there. Storyboards and outputs may live anywhere, passed as absolute paths. - Browser flows only. It drives Chromium, so it cannot capture a native app or a human's live desktop, and it cannot edit existing footage.
- Native
<select>dropdowns are drawn by the OS, outside the page, so they never appear in the recording.
Walks a project from source to a live release on Google Play's internal testing track, one verified step at a time. Handles Expo (CNG) and bare React Native.
The design principle is an inversion: authoritative where it can verify, advisory where it cannot. Anything reachable from a shell — the keystore, the signing config, the bundle's own manifest and signature — is a hard gate that blocks progress on failure. Anything inside the Play Console is a browser the agent cannot see, so those steps state a goal, offer a dated hint, and ask what the user actually sees. When the report disagrees with the hint, the skill looks up Google's current documentation instead of improvising a click path from memory.
That is what keeps it useful after Google's next redesign. A skill asserting UI it cannot observe becomes confidently wrong; one verifying artifacts it can read does not.
cd publishing-android-to-play
brew install bundletool # aapt2 cannot read an AAB
./verify.sh /path/to/app-release.aab /path/to/project/.play-release.mdGates cover: artifact is an AAB not an APK, signed with a release key rather than the debug one,
versionCode genuinely ahead of what is live, targetSdk meeting Play's floor, package match, and
a permission diff against the previous release — the check that catches a dependency quietly adding
one.
Requirements and constraints
bundletool(brew install bundletool) and a JDK forjarsigner/keytool.- Never drives a browser and never touches the Play API. The human stays in the Console deliberately; the skill's job is to make sure what reaches it is correct.
- Never stores a password. The state file holds paths and version numbers only. Credentials
live in
~/.gradle/gradle.properties, outside every repo. - Expo and bare React Native only. Flutter and native Gradle would each be one reference file — the gates and the Console half are already framework-agnostic — but shipping untested build paths as confident instructions is the exact failure this skill is built to avoid.
- Internal testing track. Not production rollouts, staged releases, or the App Store.
- Create a directory named after the skill, in kebab-case.
- Write
SKILL.mdwithnameanddescriptionfrontmatter. Spell out the situations that should trigger it — that description is the only thing the agent sees before deciding to load the skill. - Keep dependencies, examples, and generated media inside the directory, and add a
.gitignorefor anything regenerable (node_modules/, rendered video, intermediates). - Document the failures you actually hit while building it. The value of a skill is mostly in the non-obvious things that broke.
Copy or symlink a skill directory into an agent's skills path, for example:
ln -s "$PWD/recording-web-walkthroughs" ~/.claude/skills/recording-web-walkthroughs
ln -s "$PWD/recording-web-walkthroughs" ~/.cursor/skills/recording-web-walkthroughsSymlinking keeps the installed copy in step with this repo. Skills that install dependencies need
npm install run once inside the real directory, not the link.