Researcher Skill
One file. Your AI coding agent becomes a scientist.
Drop researcher.md into Claude Code, Codex, or any agent. It will design experiments, test hypotheses, discard what fails, keep what works — 30+ experiments overnight while you sleep.
What it looks like running
Experiment 7 — Replace O(n^2) neighbor search with KD-tree
Branch: research/reduce-latency · Parent: #5 · Type: real
Hypothesis: KD-tree lookup should reduce p99 from 142ms to under 100ms Changes: swapped brute-force loop in spatial_index.py with scipy.spatial.KDTree Result: p99 = 89ms (was 142ms baseline, 118ms best) — new best Status: keep
Insight: The bottleneck was always the neighbor search, not the scoring. Experiment #3 (caching) was treating the symptom.
| # |
branch |
metric |
status |
description |
| 0 |
research/reduce-latency |
142.00 |
keep |
Baseline measurement |
| 1 |
research/reduce-latency |
139.20 |
discard |
Add response caching |
| 2 |
research/reduce-latency |
141.50 |
discard |
Connection pooling tuning |
| 3 |
research/reduce-latency |
135.80 |
discard |
LRU cache on hot path |
| 5 |
research/reduce-latency |
118.00 |
keep |
Batch DB queries |
| 7 |
research/reduce-latency |
89.00 |
keep |
KD-tree neighbor search |
Example is simulated. The skill works on any codebase — run it and share your real results.
What can I research?
Anything where you can measure or evaluate a result:
- API latency — p50, p99, throughput
- Test speed — suite runtime, parallelization strategies
- Bundle size — tree-shaking, code splitting, dependency swaps
- Prompt engineering — accuracy, cost, token usage
- Algorithm tuning — runtime complexity, memory usage
- Configuration optimization — DB settings, cache sizes, thread pools
The agent interviews you about what to optimize, sets up a lab on a git branch, then works autonomously — thinking, testing, reflecting — committing before every experiment, reverting on failure, logging everything. It forks branches to explore divergent approaches, detects when it's stuck, and keeps going until you stop it or it hits a target.
Generalizes autoresearch beyond ML: supports thought experiments, non-linear branching, qualitative metrics, convergence signals, and session resume.
All experiment history lives in an untracked .lab/ directory that survives all git operations — git manages code, .lab/ manages knowledge.
License
MIT
If this is useful, star the repo and share what you discovered.
See also
Yggdrasil – your agent keeps forgetting your architecture, constraints, and past decisions. Yggdrasil gives your repository persistent semantic memory, so each task starts with the right context instead of another giant prompt.
Instagram Engineering
Shopify Engineering