Whetstone: 20 battle-tested Claude Code skills distilled from real failures

A new open-source project called Whetstone aims to encode the hard-won lessons of real-world coding failures into reusable skills for AI coding agents. Each of the 20 skills is a short, self-contained rule set distilled from a single concrete incident, sanitized so it transfers to any codebase. The project's tagline is blunt: "Agent skills honed on real work."

The key idea is that these aren't theoretical best practices. Each skill names the sanitized case it came from — a check that passed against a stand-in, a merge that dropped a field, a green test that could never fail. The lesson is stripped to its mechanism so it applies broadly.

Each skill ends with a pre-flight checklist that "fails loudly when a box is unchecked." This is a deliberate design choice, positioning the checklist as a gate on "done" rather than a suggestion that can be waved through. For developers, this is a refreshing antidote to the hand-wavy confidence that AI agents often project.

A sample of the skills

The skills cover recurring failure modes that any developer will recognize. For instance, one skill warns against trusting a green test: "after writing a test, or before trusting a green one — make it fail on purpose once, and confirm new branches are reachable by the mocks." Another addresses the trap of reading a merge conflict: "markers show where git gave up, not the full delta; auto-merge silently drops fields."

There are also skills for verifying what a build script actually runs, guarding against "not found" claims from delegated research, and designing with parameters instead of mocking machinery. The list is practical and specific, likely to resonate with anyone who's debugged a subtle issue caused by an AI agent's overconfidence.

How it works

The skills are written in plain Markdown, one folder per skill, each containing a SKILL.md file. This makes them usable by any agent that can load a system prompt — the plugin auto-discovery is the only Claude Code-specific part. You can install via the Claude Code marketplace, or manually copy the folders.

The description in each skill carries the trigger phrases, so a skill-aware agent surfaces the right skill when the situation matches. The project is open source, and the method for creating skills is itself documented, inviting others to contribute their own scars.

Why it matters for developers

This is a pragmatic tool for anyone using AI coding agents in production. It addresses a real pain point: agents that confidently claim a fix is done when it isn't. By encoding failure lessons into checkable rules, Whetstone promises to reduce those frustrating round-trips.

The approach is also notable for its humility. Instead of claiming to make agents smarter, it makes them more honest — and gives developers a way to scrutinize their work. As AI agents become more autonomous, tools like this will be essential for keeping them accountable.