▟ Whetstone

Sharpen your coding agent.

Three skills for Claude Code. Not thirteen, not forty — three. Each one earns its context tokens; nothing else made the cut.

A whetstone is the stone you sharpen a blade on. (Also: we're a pool company, so — a wet stone. We'll see ourselves out.)

MIT · free & open · 3 skills · zero config · works in any Claude Code project

Install

npx whetstone            # install into this project's .claude/skills/
npx whetstone --global   # install into ~/.claude/skills/ (all projects)
npx whetstone --list     # list the skills without installing

That's it. Claude auto-loads each skill at session start and reaches for it when it fits — or you fire it by its trigger word. Start a new session after installing.

The three

🎯 heist — steal what works

Point it at any competitor — a name, a URL, a GitHub repo — and it runs a research sweep, then hands you a ranked "steal brief": what to lift, what to leave, and why. A decision menu, not a wall of notes.

heist cursor.com

🔨 forge — sharpen the ask

Re-engineer a vague request into a sharp prompt — task, context, constraints, output, success criteria — shown to you for approval before the agent acts. Bad prompts make bad code; forge fixes them upstream.

forge add caching to the API

🪢 whip — enforce the discipline

Seven disciplines, one skill — the senior-engineer reflexes an agent skips by default:

• Prove it or don't say it — no "done" without runtime evidence you produced; a green build is parse-success, not proof.
• A failed fix means stop — re-diagnose with fresh evidence; never ship the same guess twice. Prod-only bug? Pull the real failing artifact.
• Fix the root once — bug in 2+ places means read the whole data path, fix the shared cause, grep for siblings.
• Name the bug before you re-land reverted code — "probably collateral" is how it comes back.
• Build the minimum — does it need to exist → does the codebase / stdlib / platform already do it → can it be one line?
• Cut only where the task points — no drive-by refactors; every hunk traces to the ask.
• Measure, don't eyeball — on the second CSS guess, read the rendered pixels or the DOM, then ship one targeted fix.

It fires on its own when the moment calls for it — a "done" claim, a "still broken", a growing diff — or on the word whip.

The loop

forge the request → build it → whip keeps it honest → heist what the competition got right.

How it works

Skills are just markdown in the open agentskills.io format — the same format Claude Code reads natively. Each lives in .claude/skills/<name>/SKILL.md with a short description that tells Claude when to reach for it. No config, no wiring, no runtime.

Why only three? Every installed skill's description loads into the agent's context every session — a 40-skill pack taxes every conversation to maybe use one of them. Three sharp tools beat a drawer of dull ones.

Contributing

These skills distill engineering discipline, not any one project's specifics — PRs that sharpen the methodology or fix an edge case are welcome. New skills face a high bar: it goes in one of the three, or it doesn't go in.

License

MIT — free to use, fork, and adapt.

Built by Thinkpool — where two people and their agents share one room.