evolve
/evolve
Gets the best solution, not the first one, for code with a measurable target (speed, a benchmark score, a quality metric backed by tests) by a short propose → measure → improve loop. Use when the user wants code faster, smaller or higher-scoring and there is, or can be, a scorer; or when they ask for an AlphaEvolve- or autoresearch-style search.
- Author
- bombsdontbite ↗
- Repository
- anti-skill ↗
- Licence
- MIT
- Last repository change
- 03.11.2026
Assessment and security
- Files in this package
- 7
- Scripts
- 1
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 03.11.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
- 1 scripts in this package. Review them before installing, as Claude may execute them.
Scripts: evolve31/dedup/final.py
What it does and when to use it
Gets the best solution, not the first one, for code with a measurable target (speed, a benchmark score, a quality metric backed by tests) by a short propose → measure → improve loop. Use when the user wants code faster, smaller or higher-scoring and there is, or can be, a scorer; or when they ask for an AlphaEvolve- or autoresearch-style search.
Original author description
Gets the best solution, not the first one, for code with a measurable target (speed, a benchmark score, a quality metric backed by tests) by a short propose → measure → improve loop. Use when the user wants code faster, smaller or higher-scoring and there is, or can be, a scorer; or when they ask for an AlphaEvolve- or autoresearch-style search.
How it works
- Rules
- Loop
- Output
SKILL.md preview
--- name: evolve description: Gets the best solution, not the first one, for code with a measurable target (speed, a benchmark score, a quality metric backed by tests) by a short propose → measure → improve loop. Use when the user wants code faster, smaller or higher-scoring and there is, or can be, a scorer; or when they ask for an AlphaEvolve- or autoresearch-style search. --- # evolve When a result can be measured, the first correct answer is usually not the best one, and a small loop of try → measure → improve finds better. When several different approaches land on the same number, the task is at its ceiling and more search is wasted. ## Rules 1. **Measure, never eyeball.** Use the user's scorer or write one first. It fails closed: any error, timeout, NaN or broken constraint is invalid, never a score. Do not edit it, read its hidden data or game it; if you find a hole, fix it and say so. 2. **Build a fast local harness** first: an instance generator that follows the problem statement, a correctness oracle (the reference, a brute force or a constraint checker), the same time limits. Check it ranks two candidates the way the scorer does. Iterate on it; spend scorer calls on candidates it already rates well. 3. **Keep a held-out split** (other instances or seeds) and pick the final on it, not on the numbers you tuned against. Report the drop between the two. ## Loop 1. **Strong first answer.** Write what an expert would write first, make it correct, measure it. 2. **Other families.** Build 2–3 materially different approaches (another algorithm, data structure or decomposition, not new constants) far enough to measure. 3. **Ceiling check.** If the families come within noise of each other, and of any known bound, the task is at its ceiling: tune constants once, report, and stop. 4. **Improve.** Otherwise the task is not at its ceiling, so search until the budget is spent: a correct, good answer is not a reason to stop. Run rounds of 4–6 children from the 2–3 best: mostly targeted edits, sometimes a rewrite or a crossover; drop near-duplicates before measuring. After two rounds without a gain, change the kind of change (profile and attack the hottest part, or a new family) rather than stopping. 5. **Finish.** Tune constants, then score the final pick on the held-out split. ## Output Baseline, first answer and final pick, each with search and held-out scores; the final code and in plain words why it is better; which families were tried; whether the ceiling check stopped the search early.
Files in this package (7)
- README.md 1 KB
- SKILL.md 3 KB
- agent-costs.json 1 KB
- dedup_repeats.json 1 KB
- evolve31/dedup/TASK.md 1 KB
- evolve31/dedup/calls.log 1 KB
- evolve31/dedup/final.py 2 KB
Versions
| Repository | Score | Licence |
|---|---|---|
| This page · bombsdontbite/anti-skill | 85 | MIT |
| Kanevry/session-orchestrator | 100 | MIT |
| DiogoHSM/driver | 100 | MIT |
| mycelium-hq/ai-brain-starter | 100 | MIT |
| PantherHale/qurn-skills | 100 | MIT |
| uhstray-io/huhhb | 100 | MIT |
| rikyz90/shibaclaw | 95 | Apache-2.0 |
| bombsdontbite/anti-skill | 95 | MIT |
| bombsdontbite/anti-skill | 95 | MIT |
Database snapshot: 13.11.2026 · package version eac659f3c716 (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 100autoresearch
Autonomous experiment loop: edit code, commit, run benchmark, extract metrics, keep improvements or revert, repeat forever. Use this skill when the user asks to "run autoresearch", "start an experiment loop", "optimize a metric autonomously", "autonomous experiments", "benchmark… - 100autoresearch-agent
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… - 100autoresearch-agent
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… - 100autoresearch-agent
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… - 100perf-loop
Profile and optimize a metric under its budget in a disciplined profile-optimize-reprofile loop: benchmark the whole measured set under a fixed environment, find the single highest-leverage hotspot or regression, optimize one thing, re-benchmark the WHOLE set, and repeat until… - 100self-improve
Autonomous evolutionary code improvement engine with tournament selection. Activate when user says: self-improve, self improve, evolve code, improve iteratively, tournament, benchmark loop, optimize code.