improve-skill

Runs a retrospective on a skill execution and rewrites the skill to eliminate observed friction.

5|Updated Sep 24, 2017
One-click install
npx skills add https://github.com/madyankin/dotfiles --skill improve-skill-madyankin
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: improve-skill
Source: https://github.com/madyankin/dotfiles/tree/main/.config/agents/skills/improve-skill
Command: npx skills add https://github.com/madyankin/dotfiles --skill improve-skill-madyankin

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Skills accumulate friction over repeated runs: improvised decisions, manual work that should be scripted, and instructions that get ignored. This Skill harvests that friction from a recent run and turns it into concrete, verified edits to the skill files. ## Core Features & Use Cases - Friction Harvesting: Systematically identifies improvisation, manual labor, user corrections, and dead instructions from a skill run. - Fix Classification: Maps each friction item to the right fix type, preferring shipped scripts and construction-time guarantees over checklist additions. - Closed-Loop Verification: Re-runs the edited skill to prove each change works, and records the outcome in a dated changelog. - Use Case: After a skill run where the agent hand-wrote a helper and guessed at a naming rule, invoke this Skill to ship the helper in the skill's library, codify the naming rule, and re-run to confirm both fixes hold. ## Quick Start Ask the agent to run a retrospective on the skill you just used and improve it based on what went wrong in that run.

Frequently Asked Questions about improve-skill

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I improve a skill after a bad run?▼

Invoke the retrospective immediately after the run while the context is still available. The process harvests friction from improvisation, manual labor, corrections, and dead text, then edits the skill files and re-runs the skill to prove the fixes work.

What kinds of friction does a skill retrospective look for?▼

It looks for four sources: improvisation where the agent decided something the skill did not cover, manual labor that a script could replace, user corrections indicating ambiguous instructions, and dead text that the run ignored or could not follow.

Where are user skills stored for editing?▼

User skills live in ~/.config/agents/skills/<name>/, which is the canonical path. The ~/.claude/skills path is a symlink to it, so edits through either location modify the same files.

Why is re-running the skill required after editing it?▼

An edit that has not been exercised is only a guess. Re-running on the same input confirms each change did what it was intended to do, and edge cases not hit by the re-run should be tested directly with a throwaway call.

When should friction not be fixed in a skill?▼

One-off situations, deliberate user preferences, and fixes that cost more clarity than they buy should be left alone. The report should state what was deliberately not fixed and why.