autoimprove-skill

Autonomously improves agent skills through eval-driven experimentation. Runs a Karpathy-style autoresearch loop: evaluates Recall (does the skill activate when it should?) and Precision (does it produce correct output?), proposes targeted edits, keeps improvements, and reverts regressions.

Install

npx autoimprove-skill

You'll be asked where to install:

[g] Global — all agent platforms (~/.agents, ~/.claude, ~/.opencode)
[l] Local  — this project only (./.agents/skills/autoimprove-skill)
[b] Both   — global + local

Or skip the prompt with flags:

npx autoimprove-skill --global     # global only
npx autoimprove-skill --local      # local only
npx autoimprove-skill --both       # global + local
npx autoimprove-skill --target ~/projects/my-app   # custom path

Options

| Flag | Description | |------|-------------| | -g, --global | Install globally across all agent platforms | | -l, --local | Install into the current project only | | -b, --both | Install both globally and locally | | --target <path> | Install into a specific path | | --dry-run | Show what would be installed without writing files | | --force | Overwrite existing skill files | | --uninstall | Remove the skill from all directories |

Usage

Once installed, invoke the skill by name on any supported platform:

Load the autoimprove-skill. Target: my-copywriting-skill

The skill will:

1. Locate the target skill and its evals.json
2. Ask you to choose an iteration mandate (NEVER STOP, iteration cap, stagnation cap)
3. Run the Karpathy loop — evaluate, edit, re-evaluate, keep or revert
4. Generate an HTML report when done

Eval Format

Target skills need an evals/evals.json defining test prompts. See EVALS-FORMAT.md for the full schema.

What's Inside

skill/
├── SKILL.md                  # The skill body — 5-phase workflow
├── EVALS-FORMAT.md           # evals.json schema reference
├── REPORT-TEMPLATE.html      # HTML report scaffold
└── scripts/
    └── generate-report.sh    # Converts results.tsv to HTML