Skillmint: Turn repeated work into tested skills

You correct the same step for the second time this week, and your agent still forgets it next session. Repeated workflows run by hand each time, and nothing gets written down. Skillmint turns that repetition into files your agent reads: a note, a preference, an override or a full skill that has to beat a run without it.

What it does

• Evidence first names the repeating pattern with at least two concrete examples and asks "Save this as a skill?" before writing anything.
• Route by lifespan picks the smallest home: session note, one-line preference, library entry, override for an existing skill, or a new skill.
• Draft from real work writes .agents/skills/my-<name>/SKILL.md with your trigger phrases, exact steps and one pitfall per real incident.
• Manual A/B test runs 2 or 3 realistic prompts plus one that should not trigger, with and without the skill, and keeps the skill only if it wins.
• Version and log bumps the version on every change and appends a dated line to learned.md.
• Vet outside skills checks any marketplace or repo skill for hooks, network calls, auto-commit and credential access before install.

How it works

1. Spot the repeat and show the evidence.
2. Route it to the smallest fitting home.
3. Draft the skill (under about 120 lines) and run the kit's skill linter.
4. Test by hand in fresh sessions and record results in evals.md.
5. Register, version, and retire skills on purpose when they stop earning their place.

Frameworks & tools

Works with agents that read SKILL.md files. Uses a skill linter (node scripts/lint-skills.mjs or the project's equivalent) and git.

Why this beats prompting it yourself

Asking an agent to "remember this" usually ends in a vague note or nothing. Skillmint only claims something was learned when a file was written, prefers levelling up an existing skill over a near-duplicate, and keeps a skill only after a side-by-side test.

Use cases

• Weekly routines such as deploy, migrate, seed and smoke test.
• Repeated corrections like "use pnpm, not npm" turned into an override.
• Skill maintenance to level up, split or retire personal skills.

Known limitations

The A/B test is run by hand, not automated. One-off tasks and single taste preferences are applied, not saved as skills. Deep security review of outside skills is a separate skill.