Adw Eval — Development skill for Claude Code
Measure the loop itself, then PROPOSE (never silently apply) up to 3 improvements.
How to install Adw Eval
Installs to ~/.claude/commands/frostyfruit-adw-kit-adw-eval.md
mkdir -p ~/.claude/commands && curl -fsSL https://raw.githubusercontent.com/FrostyFruit/adw-kit/HEAD/.claude/commands/adw-eval.md -o ~/.claude/commands/frostyfruit-adw-kit-adw-eval.md Restart Claude Code, or start a new session, for it to be picked up.
What Adw Eval does
description: Measure the loop itself, then PROPOSE (never silently apply) up to 3 improvements
Single job: measure how the loop is performing across all runs, find the problems that keep recurring, and propose at most THREE changes to the workflow. You do NOT edit the ADW commands in the same session that runs the eval.
Why this exists
The loop improves the product. Nothing measures the loop unless you do. And the person (or agent) running the loop cannot easily see its own blind sp
Alternatives in Development
- Evo — turns your codebase into an autoresearch loop — discovers what to measure, instruments the benchmark, then run 1.4k ★
- Hippo Feature — Build one hippo feature from RESEARCH.md using the micro-eval TDD loop 734 ★
- Overnight PRs — Autonomous overnight loop — verify issues against the codebase, close stale ones, ship PRs through /ship (bug 324 ★
Full documentation available on GitHub
View Source RepositoryRelated Skills
Ranking Eval Loop
Ranking eval loop — mine real misses, then measure the fix
Improve Skill
Propose and apply targeted improvements to a specific Claude Code skill (self-improving skills AMEND layer)
Skills Update
Apply an APPROVED improvement from the queue, with preview and apply modes. Explains each change, shows affect
Eval Cases
Grow an eval case set from the placeholder rows to a real 15-25 - harvest genuine cases out of the repository
Daily Review
Once per day, read the last 24h of state and propose 1-3 concrete improvements. Don't just report — recommend.
Autoskill
Analyze the current session for patterns and propose skill/agent improvements
Related Agents
Config Manager
Owns the global Claude Code setup and its self-improvement loop. Use via the /update-config skill (or directly
Rnd
Use this agent for RESEARCH & DEVELOPMENT — scanning the MCP ecosystem, competitor directories, upstream stack
Dotfile Drift Reporter
Detect drift between the chezmoi source state and the live $HOME files. Summarize what's out of sync and propo