I just put five skills up on Agensi, and they’re all free. A skill is a set of instructions, sometimes with a script, that your agent loads when the job calls for it. These work in Claude Code, Cursor, Copilot and most other agents that support skills.
Every one came out of my own projects, and every one exists because an agent told me something was done, tested or fine when it wasn’t. Here’s what they do:
- Tell you whether old test results still count after code or dependencies change.
- Catch editor changes that lose content on save, preview or export.
- Stop your agent at a working prototype until you’ve actually approved the direction.
- Check your live site against your style guide in a real browser.
- Playtest your browser game with a bot to find difficulty walls and unfair bosses.
1. QA Evidence Freshness Ledger
- The problem: Your agent’s tests passed yesterday. Since then the lockfile, the browser version or an uncommitted file changed. Does that green check still mean anything?
- What it does: Compares what the old result actually covered with what’s changed since, and gives one answer:
FRESH,STALE,FAILED,MISSING,UNVERIFIEDorPARTIAL. - What you get: The smallest rerun you actually need, instead of trusting stale results or rerunning everything.
- Try it: “We have a passing browser result from the previous revision. Only the lockfile changed. Which old results still apply? Don’t run the suite.”
2. Editor-to-Export Regression Guard
- Best for: Email builders, page builders, document editors and anything else with a visual editor.
- The problem: The worst bugs look fine in the editor. Then a link disappears after save, or the exported HTML says something different from the preview.
- What it does: Follows one value from the user’s click through save, reload, preview and the actual downloaded export, and compares it against a known-good baseline.
- What you get:
REGRESSION,PARTIALorNO_REGRESSION_OBSERVED. It won’t claim everything matches when there’s no baseline to compare to. - Try it: “Check the image-link change against the previous revision. Saved documents must keep their links after an unrelated edit. Cover save/reload, preview and downloaded HTML.”
3. Prototype-First Review Checkpoint
- The problem: Agents go straight from “it works” to hardening, tests and polish on a direction you never signed off on.
- What it does: Stops at a working local build and hands you the URL, the account to use, click-by-click test steps and the one product question you need to answer.
- What counts as approval: Only your actual answer. Silence, “looks cool” or a vague “continue” doesn’t count, and approving the direction is never permission to ship.
- Try it: “Add the new document-card layout, prepare a local review build with test data, and stop for my approval before hardening.”
4. Style Guide Runtime Check
- The problem: Your site drifted from its style guide anyway: a 13px toast, a gray that’s almost the right gray, a button that never picked up the page font, a modal nobody checked. Reading the CSS misses most of it, because the problem is in what the browser actually renders.
- What it does: Turns your style guide into a config and opens every screen at every screen width in a real headless browser.
- What you get: Every font size, color, border, radius and font that isn’t approved, with the exact element and the closest correct value.
- No style guide yet? It can inventory everything you’re currently using so you can write one.
- Try it: “web/STYLE_GUIDE.md is our style guide. Check whether the site actually follows it and give me a report.”
5. Autoplay Balance Test
This one came straight out of my dungeon crawler. Playtesting from level 1 to 20 by hand after every balance change wasn’t going to happen, so I had Claude build a bot that plays the game like a new player.
- How it works: Your agent writes a small adapter for your game, then the bot plays through the real UI.
- What you get: The level curve, where players wipe, difficulty walls and stuck screens.
- Boss mode: Drops the party into one boss fight at each level and measures the win rate, so “is this boss too hard?” gets an actual number.
- Proof: In testing, a fresh agent found the boss I’d made unbeatable on purpose: 0 wins in 80 fights.
- Works for: Turn-based and menu-driven browser games, not real-time action games.
- Try it: “Playtest my browser game with a bot from a fresh start and tell me whether there’s a difficulty wall and where.”
Get them
All five are listed on Agensi, free. The first three are instructions only and need nothing installed. The style check and the game tester need Node.js and Playwright in your project.
Disclosure: some links in this post are affiliate links. If you sign up or buy through one, I earn a commission.