meta-harness-optimization — Claude Code skill
A skill for automated search over model harnesses: the scaffolding that decides what an agent stores, retrieves and sees while it works.
npx skills add reason-machines/trending-skills --skill meta-harness-optimization -a claude-codeWhat meta-harness-optimization does
A harness is everything wrapped around a fixed base model — memory, retrieval, prompt assembly, tool calls. Meta-Harness searches that space instead of you hand-tuning it. A proposer agent suggests a variant, an evaluator scores it on a benchmark, and the result feeds the next round.
Two reference experiments ship with it: reference_examples/text_classification for memory-system search and reference_examples/terminal_bench_2 for scaffold evolution. Dependencies are per-experiment through uv and pyproject.toml. There is no global pip step, but the uv sync and uv run commands assume uv is on your machine.
For a new domain you start from ONBOARDING.md: open it in Claude or Cursor, talk through the task, and you get a domain_spec.md covering what the harness controls, how quality is measured and what the proposer may change. From there you write your own harness.py, evaluator.py and claude_wrapper.py.
What it helps with
- pick a memory strategy for a text classifier without manual trial and error
- run scaffold evolution on Terminal-Bench 2
- describe a new domain for harness search
- compare harness variants against a numeric score
- wire up a propose-evaluate-feedback loop for an agent
Example prompts for the agent
- Set up meta-harness for my text classification project
- Run one optimization iteration in
reference_examples/text_classification - Help me fill in
domain_spec.mdfor a new domain
How to install meta-harness-optimization
1Skills CLI (Vercel)
npx skills add reason-machines/trending-skills --skill meta-harness-optimization -a claude-code2Manually: put the files in place
~/.claude/skills/meta-harness-optimization/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/meta-harness-optimization/
1Skills CLI (Vercel)
npx skills add reason-machines/trending-skills --skill meta-harness-optimization -a codex2Manually: put the files in place
.agents/skills/meta-harness-optimization/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add reason-machines/trending-skills --skill meta-harness-optimization -a cursor2Manually: put the files in place
.cursor/skills/meta-harness-optimization/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install reason-machines/trending-skills meta-harness-optimization2Manually: put the files in place
.github/skills/meta-harness-optimization/SKILL.md1Skills CLI (Vercel)
npx skills add reason-machines/trending-skills --skill meta-harness-optimization -a gemini-cli2Platform CLI
gemini skills install https://github.com/reason-machines/trending-skills --path skills/meta-harness-optimizationManually: put the files in place
.devin/skills/meta-harness-optimization/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add reason-machines/trending-skills --skill meta-harness-optimization -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about meta-harness-optimization
How do I install meta-harness-optimization in Claude Code?
Run npx skills add reason-machines/trending-skills --skill meta-harness-optimization -a claude-code. The skill lands in ~/.claude/skills/meta-harness-optimization/SKILL.md.
What does meta-harness-optimization do?
It lays out a loop for automated harness search: a proposer agent emits a variant, an evaluator scores it on a benchmark, and the score goes back into the history for the next step.
What do I need to run the reference experiments?
The uv package manager. Each experiment runs uv sync first, then uv run, with dependencies declared in its own pyproject.toml.