[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6KB9DXXMSQTYBP73MHYH77C":3,"skill-same:01M6KB9DXXMSQTYBP73MHYH77C":153},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":28,"publishedAt":49,"linkOnly":50,"markdownContent":51,"markdownBody":52,"frontMatter":53,"refName":59,"commitSha":60,"skillDirectory":61,"forksCount":62,"repositoryHtmlUrl":63,"repositoryHomepageUrl":6,"ownerHtmlUrl":64,"ownerBlogUrl":65,"ownerDescription":6,"updatedAt":66,"install":67,"resources":80},"01M6KB9DXXMSQTYBP73MHYH77C","autoresearch-agent",null,"Autonomous experiment loop that optimizes any file by a measurable metric. Inspired by Karpathy's autoresearch. The agent edits a target file, runs a fixed evaluation, keeps improvements (git commit), discards failures (git reset), and loops indefinitely. Use when: user wants to optimize code speed, reduce bundle\u002Fimage size, improve test pass rate, optimize prompts, improve content quality (headlines, copy, CTR), or run any measurable improvement loop. Requires: a target file, an evaluation command that outputs a metric, and a git repo.","afaizalam2003\u002FMaster-claude-skill","afaizalam2003","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F222019359?v=4","01M6K8F2XNT0HKBPF00JQ78SV3","Faiz Alam","01M6K8F31EBXAPER0EG8R9HTR5",3,"MIT","https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002FSKILL.md",[18,23],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000023","Skill与智能体开发","SCENARIO","使用场景",{"categoryNo":24,"categoryName":25,"taxonomyType":26,"taxonomyTypeDesc":27},"01JBSEEDCAT000000000000030","开发工程师","PROFESSION","适用职业",[29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48],"autoresearch","autoresearch agent","command","rate","reduce","improvement loop","loops","ctr","reset","optimize","karpathy","autonomous","image","loop","improve","Evaluation","git","code","run","commit","2026-11-10 10:12:30",false,"---\nname: \"autoresearch-agent\"\ndescription: \"Autonomous experiment loop that optimizes any file by a measurable metric. Inspired by Karpathy's autoresearch. The agent edits a target file, runs a fixed evaluation, keeps improvements (git commit), discards failures (git reset), and loops indefinitely. Use when: user wants to optimize code speed, reduce bundle\u002Fimage size, improve test pass rate, optimize prompts, improve content quality (headlines, copy, CTR), or run any measurable improvement loop. Requires: a target file, an evaluation command that outputs a metric, and a git repo.\"\nlicense: MIT\nmetadata:\n  version: 2.0.0\n  author: Md Faiz Alam\n  category: engineering\n  updated: 2026-03-13\n---\n\n# Autoresearch Agent\n\n> You sleep. The agent experiments. You wake up to results.\n\nAutonomous experiment loop inspired by [Karpathy's autoresearch](https:\u002F\u002Fgithub.com\u002Fkarpathy\u002Fautoresearch). The agent edits one file, runs a fixed evaluation, keeps improvements, discards failures, and loops indefinitely.\n\nNot one guess — fifty measured attempts, compounding.\n\n---\n\n## Slash Commands\n\n| Command | What it does |\n|---------|-------------|\n| `\u002Far:setup` | Set up a new experiment interactively |\n| `\u002Far:run` | Run a single experiment iteration |\n| `\u002Far:loop` | Start autonomous loop with configurable interval (10m, 1h, daily, weekly, monthly) |\n| `\u002Far:status` | Show dashboard and results |\n| `\u002Far:resume` | Resume a paused experiment |\n\n---\n\n## When This Skill Activates\n\nRecognize these patterns from the user:\n\n- \"Make this faster \u002F smaller \u002F better\"\n- \"Optimize [file] for [metric]\"\n- \"Improve my [headlines \u002F copy \u002F prompts]\"\n- \"Run experiments overnight\"\n- \"I want to get [metric] from X to Y\"\n- Any request involving: optimize, benchmark, improve, experiment loop, autoresearch\n\nIf the user describes a target file + a way to measure success → this skill applies.\n\n---\n\n## Setup\n\n### First Time — Create the Experiment\n\nRun the setup script. The user decides where experiments live:\n\n**Project-level** (inside repo, git-tracked, shareable with team):\n```bash\npython scripts\u002Fsetup_experiment.py \\\n  --domain engineering \\\n  --name api-speed \\\n  --target src\u002Fapi\u002Fsearch.py \\\n  --eval \"pytest bench.py --tb=no -q\" \\\n  --metric p50_ms \\\n  --direction lower \\\n  --scope project\n```\n\n**User-level** (personal, in `~\u002F.autoresearch\u002F`):\n```bash\npython scripts\u002Fsetup_experiment.py \\\n  --domain marketing \\\n  --name medium-ctr \\\n  --target content\u002Ftitles.md \\\n  --eval \"python evaluate.py\" \\\n  --metric ctr_score \\\n  --direction higher \\\n  --evaluator llm_judge_content \\\n  --scope user\n```\n\nThe `--scope` flag determines where `.autoresearch\u002F` lives:\n- `project` (default) → `.autoresearch\u002F` in the repo root. Experiment definitions are git-tracked. Results are gitignored.\n- `user` → `~\u002F.autoresearch\u002F` in the home directory. Everything is personal.\n\n### What Setup Creates\n\n```\n.autoresearch\u002F\n├── config.yaml                        ← Global settings\n├── .gitignore                         ← Ignores results.tsv, *.log\n└── {domain}\u002F{experiment-name}\u002F\n    ├── program.md                     ← Objectives, constraints, strategy\n    ├── config.cfg                     ← Target, eval cmd, metric, direction\n    ├── results.tsv                    ← Experiment log (gitignored)\n    └── evaluate.py                    ← Evaluation script (if --evaluator used)\n```\n\n**results.tsv columns:** `commit | metric | status | description`\n- `commit` — short git hash\n- `metric` — float value or \"N\u002FA\" for crashes\n- `status` — keep | discard | crash\n- `description` — what changed or why it crashed\n\n### Domains\n\n| Domain | Use Cases |\n|--------|-----------|\n| `engineering` | Code speed, memory, bundle size, test pass rate, build time |\n| `marketing` | Headlines, social copy, email subjects, ad copy, engagement |\n| `content` | Article structure, SEO descriptions, readability, CTR |\n| `prompts` | System prompts, chatbot tone, agent instructions |\n| `custom` | Anything else with a measurable metric |\n\n### If `program.md` Already Exists\n\nThe user may have written their own `program.md`. If found in the experiment directory, read it. It overrides the template. Only ask for what's missing.\n\n---\n\n## Agent Protocol\n\nYou are the loop. The scripts handle setup and evaluation — you handle the creative work.\n\n### Before Starting\n1. Read `.autoresearch\u002F{domain}\u002F{name}\u002Fconfig.cfg` to get:\n   - `target` — the file you edit\n   - `evaluate_cmd` — the command that measures your changes\n   - `metric` — the metric name to look for in eval output\n   - `metric_direction` — \"lower\" or \"higher\" is better\n   - `time_budget_minutes` — max time per evaluation\n2. Read `program.md` for strategy, constraints, and what you can\u002Fcannot change\n3. Read `results.tsv` for experiment history (columns: commit, metric, status, description)\n4. Checkout the experiment branch: `git checkout autoresearch\u002F{domain}\u002F{name}`\n\n### Each Iteration\n1. Review results.tsv — what worked? What failed? What hasn't been tried?\n2. Decide ONE change to the target file. One variable per experiment.\n3. Edit the target file\n4. Commit: `git add {target} && git commit -m \"experiment: {description}\"`\n5. Evaluate: `python scripts\u002Frun_experiment.py --experiment {domain}\u002F{name} --single`\n6. Read the output — it prints KEEP, DISCARD, or CRASH with the metric value\n7. Go to step 1\n\n### What the Script Handles (you don't)\n- Running the eval command with timeout\n- Parsing the metric from eval output\n- Comparing to previous best\n- Reverting the commit on failure (`git reset --hard HEAD~1`)\n- Logging the result to results.tsv\n\n### Starting an Experiment\n\n```bash\n# Single iteration (the agent calls this repeatedly)\npython scripts\u002Frun_experiment.py --experiment engineering\u002Fapi-speed --single\n\n# Dry run (test setup before starting)\npython scripts\u002Frun_experiment.py --experiment engineering\u002Fapi-speed --dry-run\n```\n\n### Strategy Escalation\n- Runs 1-5: Low-hanging fruit (obvious improvements, simple optimizations)\n- Runs 6-15: Systematic exploration (vary one parameter at a time)\n- Runs 16-30: Structural changes (algorithm swaps, architecture shifts)\n- Runs 30+: Radical experiments (completely different approaches)\n- If no improvement in 20+ runs: update program.md Strategy section\n\n### Self-Improvement\nAfter every 10 experiments, review results.tsv for patterns. Update the\nStrategy section of program.md with what you learned (e.g., \"caching changes\nconsistently improve by 5-10%\", \"refactoring attempts never improve the metric\").\nFuture iterations benefit from this accumulated knowledge.\n\n### Stopping\n- Run until interrupted by the user, context limit reached, or goal in program.md is met\n- Before stopping: ensure results.tsv is up to date\n- On context limit: the next session can resume — results.tsv and git log persist\n\n### Rules\n\n- **One change per experiment.** Don't change 5 things at once. You won't know what worked.\n- **Simplicity criterion.** A small improvement that adds ugly complexity is not worth it. Equal performance with simpler code is a win. Removing code that gets same results is the best outcome.\n- **Never modify the evaluator.** `evaluate.py` is the ground truth. Modifying it invalidates all comparisons. Hard stop if you catch yourself doing this.\n- **Timeout.** If a run exceeds 2.5× the time budget, kill it and treat as crash.\n- **Crash handling.** If it's a typo or missing import, fix and re-run. If the idea is fundamentally broken, revert, log \"crash\", move on. 5 consecutive crashes → pause and alert.\n- **No new dependencies.** Only use what's already available in the project.\n\n---\n\n## Evaluators\n\nReady-to-use evaluation scripts. Copied into the experiment directory during setup with `--evaluator`.\n\n### Free Evaluators (no API cost)\n\n| Evaluator | Metric | Use Case |\n|-----------|--------|----------|\n| `benchmark_speed` | `p50_ms` (lower) | Function\u002FAPI execution time |\n| `benchmark_size` | `size_bytes` (lower) | File, bundle, Docker image size |\n| `test_pass_rate` | `pass_rate` (higher) | Test suite pass percentage |\n| `build_speed` | `build_seconds` (lower) | Build\u002Fcompile\u002FDocker build time |\n| `memory_usage` | `peak_mb` (lower) | Peak memory during execution |\n\n### LLM Judge Evaluators (uses your subscription)\n\n| Evaluator | Metric | Use Case |\n|-----------|--------|----------|\n| `llm_judge_content` | `ctr_score` 0-10 (higher) | Headlines, titles, descriptions |\n| `llm_judge_prompt` | `quality_score` 0-100 (higher) | System prompts, agent instructions |\n| `llm_judge_copy` | `engagement_score` 0-10 (higher) | Social posts, ad copy, emails |\n\nLLM judges call the CLI tool the user is already running (Claude, Codex, Gemini). The evaluation prompt is locked inside `evaluate.py` — the agent cannot modify it. This prevents the agent from gaming its own evaluator.\n\nThe user's existing subscription covers the cost:\n- Claude Code Max → unlimited Claude calls for evaluation\n- Codex CLI (ChatGPT Pro) → unlimited Codex calls\n- Gemini CLI (free tier) → free evaluation calls\n\n### Custom Evaluators\n\nIf no built-in evaluator fits, the user writes their own `evaluate.py`. Only requirement: it must print `metric_name: value` to stdout.\n\n```python\n#!\u002Fusr\u002Fbin\u002Fenv python3\n# My custom evaluator — DO NOT MODIFY after experiment starts\nimport subprocess\nresult = subprocess.run([\"my-benchmark\", \"--json\"], capture_output=True, text=True)\n# Parse and output\nprint(f\"my_metric: {parse_score(result.stdout)}\")\n```\n\n---\n\n## Viewing Results\n\n```bash\n# Single experiment\npython scripts\u002Flog_results.py --experiment engineering\u002Fapi-speed\n\n# All experiments in a domain\npython scripts\u002Flog_results.py --domain engineering\n\n# Cross-experiment dashboard\npython scripts\u002Flog_results.py --dashboard\n\n# Export formats\npython scripts\u002Flog_results.py --experiment engineering\u002Fapi-speed --format csv --output results.csv\npython scripts\u002Flog_results.py --experiment engineering\u002Fapi-speed --format markdown --output results.md\npython scripts\u002Flog_results.py --dashboard --format markdown --output dashboard.md\n```\n\n### Dashboard Output\n\n```\nDOMAIN          EXPERIMENT          RUNS  KEPT  BEST         Δ FROM START  STATUS\nengineering     api-speed            47    14   185ms        -76.9%        active\nengineering     bundle-size          23     8   412KB        -58.3%        paused\nmarketing       medium-ctr           31    11   8.4\u002F10       +68.0%        active\nprompts         support-tone         15     6   82\u002F100       +46.4%        done\n```\n\n### Export Formats\n\n- **TSV** — default, tab-separated (compatible with spreadsheets)\n- **CSV** — comma-separated, with proper quoting\n- **Markdown** — formatted table, readable in GitHub\u002Fdocs\n\n---\n\n## Proactive Triggers\n\nFlag these without being asked:\n\n- **No evaluation command works** → Test it before starting the loop. Run once, verify output.\n- **Target file not in git** → `git init && git add . && git commit -m 'initial'` first.\n- **Metric direction unclear** → Ask: is lower or higher better? Must know before starting.\n- **Time budget too short** → If eval takes longer than budget, every run crashes.\n- **Agent modifying evaluate.py** → Hard stop. This invalidates all comparisons.\n- **5 consecutive crashes** → Pause the loop. Alert the user. Don't keep burning cycles.\n- **No improvement in 20+ runs** → Suggest changing strategy in program.md or trying a different approach.\n\n---\n\n## Installation\n\n### One-liner (any tool)\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill.git\ncp -r claude-skills\u002Fengineering\u002Fautoresearch-agent ~\u002F.claude\u002Fskills\u002F\n```\n\n### Multi-tool install\n```bash\n.\u002Fscripts\u002Fconvert.sh --skill autoresearch-agent --tool codex|gemini|cursor|windsurf|openclaw\n```\n\n### OpenClaw\n```bash\nclawhub install cs-autoresearch-agent\n```\n\n---\n\n## Related Skills\n\n- **self-improving-agent** — improves an agent's own memory\u002Frules over time. NOT for structured experiment loops.\n- **senior-ml-engineer** — ML architecture decisions. Complementary — use for initial design, then autoresearch for optimization.\n- **tdd-guide** — test-driven development. Complementary — tests can be the evaluation function.\n- **skill-security-auditor** — audit skills before publishing. NOT for optimization loops.\n","# Autoresearch Agent\n\n> You sleep. The agent experiments. You wake up to results.\n\nAutonomous experiment loop inspired by [Karpathy's autoresearch](https:\u002F\u002Fgithub.com\u002Fkarpathy\u002Fautoresearch). The agent edits one file, runs a fixed evaluation, keeps improvements, discards failures, and loops indefinitely.\n\nNot one guess — fifty measured attempts, compounding.\n\n---\n\n## Slash Commands\n\n| Command | What it does |\n|---------|-------------|\n| `\u002Far:setup` | Set up a new experiment interactively |\n| `\u002Far:run` | Run a single experiment iteration |\n| `\u002Far:loop` | Start autonomous loop with configurable interval (10m, 1h, daily, weekly, monthly) |\n| `\u002Far:status` | Show dashboard and results |\n| `\u002Far:resume` | Resume a paused experiment |\n\n---\n\n## When This Skill Activates\n\nRecognize these patterns from the user:\n\n- \"Make this faster \u002F smaller \u002F better\"\n- \"Optimize [file] for [metric]\"\n- \"Improve my [headlines \u002F copy \u002F prompts]\"\n- \"Run experiments overnight\"\n- \"I want to get [metric] from X to Y\"\n- Any request involving: optimize, benchmark, improve, experiment loop, autoresearch\n\nIf the user describes a target file + a way to measure success → this skill applies.\n\n---\n\n## Setup\n\n### First Time — Create the Experiment\n\nRun the setup script. The user decides where experiments live:\n\n**Project-level** (inside repo, git-tracked, shareable with team):\n```bash\npython scripts\u002Fsetup_experiment.py \\\n  --domain engineering \\\n  --name api-speed \\\n  --target src\u002Fapi\u002Fsearch.py \\\n  --eval \"pytest bench.py --tb=no -q\" \\\n  --metric p50_ms \\\n  --direction lower \\\n  --scope project\n```\n\n**User-level** (personal, in `~\u002F.autoresearch\u002F`):\n```bash\npython scripts\u002Fsetup_experiment.py \\\n  --domain marketing \\\n  --name medium-ctr \\\n  --target content\u002Ftitles.md \\\n  --eval \"python evaluate.py\" \\\n  --metric ctr_score \\\n  --direction higher \\\n  --evaluator llm_judge_content \\\n  --scope user\n```\n\nThe `--scope` flag determines where `.autoresearch\u002F` lives:\n- `project` (default) → `.autoresearch\u002F` in the repo root. Experiment definitions are git-tracked. Results are gitignored.\n- `user` → `~\u002F.autoresearch\u002F` in the home directory. Everything is personal.\n\n### What Setup Creates\n\n```\n.autoresearch\u002F\n├── config.yaml                        ← Global settings\n├── .gitignore                         ← Ignores results.tsv, *.log\n└── {domain}\u002F{experiment-name}\u002F\n    ├── program.md                     ← Objectives, constraints, strategy\n    ├── config.cfg                     ← Target, eval cmd, metric, direction\n    ├── results.tsv                    ← Experiment log (gitignored)\n    └── evaluate.py                    ← Evaluation script (if --evaluator used)\n```\n\n**results.tsv columns:** `commit | metric | status | description`\n- `commit` — short git hash\n- `metric` — float value or \"N\u002FA\" for crashes\n- `status` — keep | discard | crash\n- `description` — what changed or why it crashed\n\n### Domains\n\n| Domain | Use Cases |\n|--------|-----------|\n| `engineering` | Code speed, memory, bundle size, test pass rate, build time |\n| `marketing` | Headlines, social copy, email subjects, ad copy, engagement |\n| `content` | Article structure, SEO descriptions, readability, CTR |\n| `prompts` | System prompts, chatbot tone, agent instructions |\n| `custom` | Anything else with a measurable metric |\n\n### If `program.md` Already Exists\n\nThe user may have written their own `program.md`. If found in the experiment directory, read it. It overrides the template. Only ask for what's missing.\n\n---\n\n## Agent Protocol\n\nYou are the loop. The scripts handle setup and evaluation — you handle the creative work.\n\n### Before Starting\n1. Read `.autoresearch\u002F{domain}\u002F{name}\u002Fconfig.cfg` to get:\n   - `target` — the file you edit\n   - `evaluate_cmd` — the command that measures your changes\n   - `metric` — the metric name to look for in eval output\n   - `metric_direction` — \"lower\" or \"higher\" is better\n   - `time_budget_minutes` — max time per evaluation\n2. Read `program.md` for strategy, constraints, and what you can\u002Fcannot change\n3. Read `results.tsv` for experiment history (columns: commit, metric, status, description)\n4. Checkout the experiment branch: `git checkout autoresearch\u002F{domain}\u002F{name}`\n\n### Each Iteration\n1. Review results.tsv — what worked? What failed? What hasn't been tried?\n2. Decide ONE change to the target file. One variable per experiment.\n3. Edit the target file\n4. Commit: `git add {target} && git commit -m \"experiment: {description}\"`\n5. Evaluate: `python scripts\u002Frun_experiment.py --experiment {domain}\u002F{name} --single`\n6. Read the output — it prints KEEP, DISCARD, or CRASH with the metric value\n7. Go to step 1\n\n### What the Script Handles (you don't)\n- Running the eval command with timeout\n- Parsing the metric from eval output\n- Comparing to previous best\n- Reverting the commit on failure (`git reset --hard HEAD~1`)\n- Logging the result to results.tsv\n\n### Starting an Experiment\n\n```bash\n# Single iteration (the agent calls this repeatedly)\npython scripts\u002Frun_experiment.py --experiment engineering\u002Fapi-speed --single\n\n# Dry run (test setup before starting)\npython scripts\u002Frun_experiment.py --experiment engineering\u002Fapi-speed --dry-run\n```\n\n### Strategy Escalation\n- Runs 1-5: Low-hanging fruit (obvious improvements, simple optimizations)\n- Runs 6-15: Systematic exploration (vary one parameter at a time)\n- Runs 16-30: Structural changes (algorithm swaps, architecture shifts)\n- Runs 30+: Radical experiments (completely different approaches)\n- If no improvement in 20+ runs: update program.md Strategy section\n\n### Self-Improvement\nAfter every 10 experiments, review results.tsv for patterns. Update the\nStrategy section of program.md with what you learned (e.g., \"caching changes\nconsistently improve by 5-10%\", \"refactoring attempts never improve the metric\").\nFuture iterations benefit from this accumulated knowledge.\n\n### Stopping\n- Run until interrupted by the user, context limit reached, or goal in program.md is met\n- Before stopping: ensure results.tsv is up to date\n- On context limit: the next session can resume — results.tsv and git log persist\n\n### Rules\n\n- **One change per experiment.** Don't change 5 things at once. You won't know what worked.\n- **Simplicity criterion.** A small improvement that adds ugly complexity is not worth it. Equal performance with simpler code is a win. Removing code that gets same results is the best outcome.\n- **Never modify the evaluator.** `evaluate.py` is the ground truth. Modifying it invalidates all comparisons. Hard stop if you catch yourself doing this.\n- **Timeout.** If a run exceeds 2.5× the time budget, kill it and treat as crash.\n- **Crash handling.** If it's a typo or missing import, fix and re-run. If the idea is fundamentally broken, revert, log \"crash\", move on. 5 consecutive crashes → pause and alert.\n- **No new dependencies.** Only use what's already available in the project.\n\n---\n\n## Evaluators\n\nReady-to-use evaluation scripts. Copied into the experiment directory during setup with `--evaluator`.\n\n### Free Evaluators (no API cost)\n\n| Evaluator | Metric | Use Case |\n|-----------|--------|----------|\n| `benchmark_speed` | `p50_ms` (lower) | Function\u002FAPI execution time |\n| `benchmark_size` | `size_bytes` (lower) | File, bundle, Docker image size |\n| `test_pass_rate` | `pass_rate` (higher) | Test suite pass percentage |\n| `build_speed` | `build_seconds` (lower) | Build\u002Fcompile\u002FDocker build time |\n| `memory_usage` | `peak_mb` (lower) | Peak memory during execution |\n\n### LLM Judge Evaluators (uses your subscription)\n\n| Evaluator | Metric | Use Case |\n|-----------|--------|----------|\n| `llm_judge_content` | `ctr_score` 0-10 (higher) | Headlines, titles, descriptions |\n| `llm_judge_prompt` | `quality_score` 0-100 (higher) | System prompts, agent instructions |\n| `llm_judge_copy` | `engagement_score` 0-10 (higher) | Social posts, ad copy, emails |\n\nLLM judges call the CLI tool the user is already running (Claude, Codex, Gemini). The evaluation prompt is locked inside `evaluate.py` — the agent cannot modify it. This prevents the agent from gaming its own evaluator.\n\nThe user's existing subscription covers the cost:\n- Claude Code Max → unlimited Claude calls for evaluation\n- Codex CLI (ChatGPT Pro) → unlimited Codex calls\n- Gemini CLI (free tier) → free evaluation calls\n\n### Custom Evaluators\n\nIf no built-in evaluator fits, the user writes their own `evaluate.py`. Only requirement: it must print `metric_name: value` to stdout.\n\n```python\n#!\u002Fusr\u002Fbin\u002Fenv python3\n# My custom evaluator — DO NOT MODIFY after experiment starts\nimport subprocess\nresult = subprocess.run([\"my-benchmark\", \"--json\"], capture_output=True, text=True)\n# Parse and output\nprint(f\"my_metric: {parse_score(result.stdout)}\")\n```\n\n---\n\n## Viewing Results\n\n```bash\n# Single experiment\npython scripts\u002Flog_results.py --experiment engineering\u002Fapi-speed\n\n# All experiments in a domain\npython scripts\u002Flog_results.py --domain engineering\n\n# Cross-experiment dashboard\npython scripts\u002Flog_results.py --dashboard\n\n# Export formats\npython scripts\u002Flog_results.py --experiment engineering\u002Fapi-speed --format csv --output results.csv\npython scripts\u002Flog_results.py --experiment engineering\u002Fapi-speed --format markdown --output results.md\npython scripts\u002Flog_results.py --dashboard --format markdown --output dashboard.md\n```\n\n### Dashboard Output\n\n```\nDOMAIN          EXPERIMENT          RUNS  KEPT  BEST         Δ FROM START  STATUS\nengineering     api-speed            47    14   185ms        -76.9%        active\nengineering     bundle-size          23     8   412KB        -58.3%        paused\nmarketing       medium-ctr           31    11   8.4\u002F10       +68.0%        active\nprompts         support-tone         15     6   82\u002F100       +46.4%        done\n```\n\n### Export Formats\n\n- **TSV** — default, tab-separated (compatible with spreadsheets)\n- **CSV** — comma-separated, with proper quoting\n- **Markdown** — formatted table, readable in GitHub\u002Fdocs\n\n---\n\n## Proactive Triggers\n\nFlag these without being asked:\n\n- **No evaluation command works** → Test it before starting the loop. Run once, verify output.\n- **Target file not in git** → `git init && git add . && git commit -m 'initial'` first.\n- **Metric direction unclear** → Ask: is lower or higher better? Must know before starting.\n- **Time budget too short** → If eval takes longer than budget, every run crashes.\n- **Agent modifying evaluate.py** → Hard stop. This invalidates all comparisons.\n- **5 consecutive crashes** → Pause the loop. Alert the user. Don't keep burning cycles.\n- **No improvement in 20+ runs** → Suggest changing strategy in program.md or trying a different approach.\n\n---\n\n## Installation\n\n### One-liner (any tool)\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill.git\ncp -r claude-skills\u002Fengineering\u002Fautoresearch-agent ~\u002F.claude\u002Fskills\u002F\n```\n\n### Multi-tool install\n```bash\n.\u002Fscripts\u002Fconvert.sh --skill autoresearch-agent --tool codex|gemini|cursor|windsurf|openclaw\n```\n\n### OpenClaw\n```bash\nclawhub install cs-autoresearch-agent\n```\n\n---\n\n## Related Skills\n\n- **self-improving-agent** — improves an agent's own memory\u002Frules over time. NOT for structured experiment loops.\n- **senior-ml-engineer** — ML architecture decisions. Complementary — use for initial design, then autoresearch for optimization.\n- **tdd-guide** — test-driven development. Complementary — tests can be the evaluation function.\n- **skill-security-auditor** — audit skills before publishing. NOT for optimization loops.\n",{"name":5,"license":15,"metadata":54,"description":7},{"author":55,"updated":56,"version":57,"category":58},"Md Faiz Alam",1773360000000,"2.0.0","engineering","main","0b48cbb0e6d5be6ff68c0f6b28bc0c65ee21c4a1","engineering\u002Fautoresearch-agent",2,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill","https:\u002F\u002Fgithub.com\u002Fafaizalam2003","https:\u002F\u002Fwww.linkedin.com\u002Fin\u002Fmdfaizalam2003\u002F","2026-04-01 04:35:43",{"source":8,"skillName":5,"commands":68,"prompt":78,"zipAvailable":79},[69,72,75],{"tool":70,"command":71},"npx","npx skills add afaizalam2003\u002FMaster-claude-skill --skill autoresearch-agent",{"tool":73,"command":74},"bunx","bunx skills add afaizalam2003\u002FMaster-claude-skill --skill autoresearch-agent",{"tool":76,"command":77},"pnpm","pnpm dlx skills add afaizalam2003\u002FMaster-claude-skill --skill autoresearch-agent","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002FSKILL.md",true,[81,87,91,95,101,105,109,113,117,121,125,129,133,137,141,145,149],{"path":82,"resourceType":83,"resourceTypeDesc":84,"sizeBytes":85,"sourceHtmlUrl":86},".claude-plugin\u002Fplugin.json","REFERENCE","参考文档",539,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002F.claude-plugin\u002Fplugin.json",{"path":88,"resourceType":83,"resourceTypeDesc":84,"sizeBytes":89,"sourceHtmlUrl":90},"agents\u002Fexperiment-runner.md",3266,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002Fagents\u002Fexperiment-runner.md",{"path":92,"resourceType":83,"resourceTypeDesc":84,"sizeBytes":93,"sourceHtmlUrl":94},"CLAUDE.md",1850,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002FCLAUDE.md",{"path":96,"resourceType":97,"resourceTypeDesc":98,"sizeBytes":99,"sourceHtmlUrl":100},"evaluators\u002Fbenchmark_size.py","SCRIPT","脚本",1899,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002Fevaluators\u002Fbenchmark_size.py",{"path":102,"resourceType":97,"resourceTypeDesc":98,"sizeBytes":103,"sourceHtmlUrl":104},"evaluators\u002Fbenchmark_speed.py",1179,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002Fevaluators\u002Fbenchmark_speed.py",{"path":106,"resourceType":97,"resourceTypeDesc":98,"sizeBytes":107,"sourceHtmlUrl":108},"evaluators\u002Fbuild_speed.py",1141,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002Fevaluators\u002Fbuild_speed.py",{"path":110,"resourceType":97,"resourceTypeDesc":98,"sizeBytes":111,"sourceHtmlUrl":112},"evaluators\u002Fllm_judge_content.py",2714,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002Fevaluators\u002Fllm_judge_content.py",{"path":114,"resourceType":97,"resourceTypeDesc":98,"sizeBytes":115,"sourceHtmlUrl":116},"evaluators\u002Fllm_judge_copy.py",4471,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002Fevaluators\u002Fllm_judge_copy.py",{"path":118,"resourceType":97,"resourceTypeDesc":98,"sizeBytes":119,"sourceHtmlUrl":120},"evaluators\u002Fllm_judge_prompt.py",3026,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002Fevaluators\u002Fllm_judge_prompt.py",{"path":122,"resourceType":97,"resourceTypeDesc":98,"sizeBytes":123,"sourceHtmlUrl":124},"evaluators\u002Fmemory_usage.py",1663,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002Fevaluators\u002Fmemory_usage.py",{"path":126,"resourceType":97,"resourceTypeDesc":98,"sizeBytes":127,"sourceHtmlUrl":128},"evaluators\u002Ftest_pass_rate.py",1586,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002Fevaluators\u002Ftest_pass_rate.py",{"path":130,"resourceType":83,"resourceTypeDesc":84,"sizeBytes":131,"sourceHtmlUrl":132},"references\u002Fexperiment-domains.md",6672,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002Freferences\u002Fexperiment-domains.md",{"path":134,"resourceType":83,"resourceTypeDesc":84,"sizeBytes":135,"sourceHtmlUrl":136},"references\u002Fprogram-template.md",4819,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002Freferences\u002Fprogram-template.md",{"path":138,"resourceType":97,"resourceTypeDesc":98,"sizeBytes":139,"sourceHtmlUrl":140},"scripts\u002Flog_results.py",15327,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002Fscripts\u002Flog_results.py",{"path":142,"resourceType":97,"resourceTypeDesc":98,"sizeBytes":143,"sourceHtmlUrl":144},"scripts\u002Frun_experiment.py",9821,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002Fscripts\u002Frun_experiment.py",{"path":146,"resourceType":97,"resourceTypeDesc":98,"sizeBytes":147,"sourceHtmlUrl":148},"scripts\u002Fsetup_experiment.py",14341,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002Fscripts\u002Fsetup_experiment.py",{"path":150,"resourceType":83,"resourceTypeDesc":84,"sizeBytes":151,"sourceHtmlUrl":152},"settings.json",674,"https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fengineering\u002Fautoresearch-agent\u002Fsettings.json",[154,184,208,233,263,283],{"skillNo":155,"name":156,"nameZh":6,"description":157,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":158,"categories":159,"keywords":163,"publishedAt":49},"01M6KC4E8MYR9DBAZTAXHKYMMH","soc2-compliance","Use when the user asks to prepare for SOC 2 audits, map Trust Service Criteria, build control matrices, collect audit evidence, perform gap analysis, or assess SOC 2 Type I vs Type II readiness.","https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fra-qm-team\u002Fsoc2-compliance\u002FSKILL.md",[160],{"categoryNo":161,"categoryName":162,"taxonomyType":26,"taxonomyTypeDesc":27},"01JBSEEDCAT000000000000037","律师与法务",[164,165,166,167,168,169,170,171,172,173,174,175,176,177,178,179,180,181,182,183],"compliance","soc2","soc2 compliance","audit evidence","audit","collect","soc","analysis","use","evidence","trust","build","map","control","type","gap","readiness","gap analysis","assess","for",{"skillNo":185,"name":186,"nameZh":6,"description":187,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":188,"categories":189,"keywords":190,"publishedAt":49},"01M6KC461R3GKGZ1TAYHVR1MW9","risk-management-specialist","Medical device risk management specialist implementing ISO 14971 throughout product lifecycle. Provides risk analysis, risk evaluation, risk control, and post-production information analysis. Use when user mentions risk management, ISO 14971, risk analysis, FMEA, fault tree analysis, hazard identification, risk control, risk matrix, benefit-risk analysis, residual risk, risk acceptability, or post-market risk.","https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fra-qm-team\u002Frisk-management-specialist\u002FSKILL.md",[],[191,192,193,194,195,196,44,197,171,172,198,199,200,201,202,203,204,205,206,207],"risk","risk management specialist","management","risk management","risk evaluation","product","lifecycle","risk matrix","iso","medical","iso 14971","identification","information","post-production","tree","matrix","implementing",{"skillNo":209,"name":210,"nameZh":6,"description":211,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":212,"categories":213,"keywords":214,"publishedAt":49},"01M6KC3HF5ZFSSP2NZF68PEPFB","regulatory-affairs-head","Senior Regulatory Affairs Manager for HealthTech and MedTech companies. Prepares FDA 510(k), De Novo, and PMA submission packages; analyzes regulatory pathways for new medical devices; drafts responses to FDA deficiency letters and Notified Body queries; develops CE marking technical documentation under EU MDR 2017\u002F745; coordinates multi-market approval strategies across FDA, EU, Health Canada, PMDA, and NMPA; and maintains regulatory intelligence on evolving standards. Use when users need to plan or execute FDA submissions, navigate 510(k) or PMA approval processes, achieve CE marking, prepare pre-submission meeting materials, write regulatory strategy documents, respond to agency queries, or manage compliance documentation for medical device market access.","https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fra-qm-team\u002Fregulatory-affairs-head\u002FSKILL.md",[],[215,216,217,218,219,220,221,222,223,172,224,225,226,227,228,229,230,231,232,183],"regulatory affairs head","regulatory","head","documents","devices","health","plan","write","strategy","standards","queries","manage","strategies","de","users","respond","medical devices","manager",{"skillNo":234,"name":235,"nameZh":6,"description":236,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":237,"categories":238,"keywords":245,"publishedAt":49},"01M6KC34Z8M1559EY1CXYBJB1A","quality-manager-qms-iso13485","ISO 13485 Quality Management System implementation and maintenance for medical device organizations. Provides QMS design, documentation control, internal auditing, CAPA management, and certification support. Use when working with medical device quality systems, preparing for ISO 13485 audits, managing regulatory compliance documentation, setting up corrective actions, or building audit preparation programs. Useful for quality management, audit preparation, regulatory compliance, medical device documentation, and corrective action workflows.","https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fra-qm-team\u002Fquality-manager-qms-iso13485\u002FSKILL.md",[239,242],{"categoryNo":240,"categoryName":241,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000015","视觉设计",{"categoryNo":243,"categoryName":244,"taxonomyType":26,"taxonomyTypeDesc":27},"01JBSEEDCAT000000000000032","设计师",[246,247,232,248,249,250,251,252,253,172,254,255,256,257,258,259,260,261,183,262],"quality manager qms iso13485","quality","design","workflows","validation","actions","certification","system","maintenance","document","systems","records","ISO 13485","organizations","corrective action","internal audit","preparation",{"skillNo":264,"name":265,"nameZh":6,"description":266,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":267,"categories":268,"keywords":270,"publishedAt":49},"01M6KC2FXHBHXYX2ZZ17Y3GDB4","quality-manager-qmr","Senior Quality Manager Responsible Person (QMR) for HealthTech and MedTech companies. Provides quality system governance, management review leadership, regulatory compliance oversight, and quality performance monitoring per ISO 13485 Clause 5.5.2.","https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fra-qm-team\u002Fquality-manager-qmr\u002FSKILL.md",[269],{"categoryNo":161,"categoryName":162,"taxonomyType":26,"taxonomyTypeDesc":27},[271,247,232,272,273,274,275,276,164,277,253,278,216,279,280,281,199,258,183,282],"quality manager qmr","review","policy","performance","governance","monitoring","regulatory compliance","oversight","cost","culture","leadership","person",{"skillNo":284,"name":285,"nameZh":6,"description":286,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":287,"categories":288,"keywords":294,"publishedAt":49},"01M6KC1VRR5JTDBNZANHPS6VR6","quality-documentation-manager","Document control system management for medical device QMS. Covers document numbering, version control, change management, and 21 CFR Part 11 compliance. Use for document control procedures, change control workflow, document numbering, version management, electronic signature compliance, or regulatory documentation review.","https:\u002F\u002Fgithub.com\u002Fafaizalam2003\u002FMaster-claude-skill\u002Fblob\u002Fmain\u002Fra-qm-team\u002Fquality-documentation-manager\u002FSKILL.md",[289,292,293],{"categoryNo":290,"categoryName":291,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000002","文档处理",{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},{"categoryNo":161,"categoryName":162,"taxonomyType":26,"taxonomyTypeDesc":27},[295,296,247,232,272,297,298,168,253,299,300,301,302,303,200,304,305,306,307,183],"documentation","quality documentation manager","workflow","change","change management","list","master","system management","documentation review","cfr","21 cfr part 11","change control","retention"]