skill-scientist
Skill1 starsChanged 3 months agoMITView on GitHub
This is a thin human-facing router. Use skill.yaml as the source of truth for exact commands, inputs, outputs, and workflow chaining.
In this repository50 other files it gives its agents.
- AGENTS.mdAGENTS.md
- .claude/CLAUDE.mdCLAUDE.md
- CLAUDE.mdCLAUDE.md
- agent-safetySkill
- ai-llmSkill
- apiSkill
- asset-graphSkill
- authSkill
- auto-researchSkill
- binary-analysisSkill
- business-logicSkill
- campaignSkill
- cicd-securitySkill
- cloudSkill
- cors-csrfSkill
- coverageSkill
- cross-accountSkill
- domain-modelSkill
- enrichmentSkill
- evaluation-harnessSkill
- file-uploadSkill
- graphql-auditSkill
- http-protocolSkill
- identity-domainSkill
- impact-verifierSkill
- meme-coin-auditSkill
- mobileSkill
- modern-browserSkill
- nuclei-scannerSkill
- oob-infraSkill
- osintSkill
- personaSkill
- plannerSkill
- privescSkill
- program-memorySkill
- race-conditionSkill
- rceSkill
- reconSkill
- reportingSkill
- sastSkill
- scope-managerSkill
- sqliSkill
- ssrfSkill
- standard-catalogSkill
- technique-kbSkill
- traffic-corpusSkill
- triage-validationSkill
- vuln-intelSkill
- web3-auditSkill
- xssSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 0 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 16
Start with generate-hypothesis unless a hypothesis file already exists from a prior run.
- Very specific: line 17
Follow each workflow next mapping in skill.yaml after reviewing generated outputs.
- Very specific: line 18
If a fixture does not exist, create it manually or use the evaluation-harness skill before re-running run-experiment.
- Very specific: line 19
Do NOT run promote-skill-update without reviewing the review report scores.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Skill Scientist
Overview
AI Scientist pattern adapted for bug bounty skill improvement. End-to-end automated pipeline: Hypothesize -> Design -> Run in lab -> Review -> Promote. Generates and validates improvements to existing fat skills by identifying coverage gaps, designing controlled experiments with positive/negative controls, executing against evaluation fixtures, scoring results, and producing promotion proposals. Never modifies production skills without explicit approval.
This is a thin human-facing router. Use skill.yaml as the source of truth for exact commands, inputs, outputs, and workflow chaining.
Quick Reference
- Skill:
skill-scientist - Severity range:
info - Required tools:
python3,jq,bash - Expected input files:
.claude/skills/skill-scientist/payloads/coverage_matrix.yaml - Scope check: confirm authorization before running experiments or fixture creation.
Workflow Selection
- Start with
generate-hypothesisunless a hypothesis file already exists from a prior run. - Follow each workflow
nextmapping inskill.yamlafter reviewing generated outputs. - If a fixture does not exist, create it manually or use the evaluation-harness skill before re-running
run-experiment. - Do NOT run
promote-skill-updatewithout reviewing the review report scores.
Available Workflows
| Workflow | Purpose | Script paths | Primary outputs | Evidence |
generate-hypothesis | Read coverage matrix, find high-priority gaps, output hypotheses. | .claude/skills/skill-scientist/scripts/generate_hypothesis.py | $OUTDIR/hypotheses.json | $OUTDIR/evidence/ |
design-experiment | Design controlled experiments with positive/negative controls per hypothesis. | .claude/skills/skill-scientist/scripts/design_experiment.py | $OUTDIR/design_manifest.json | $OUTDIR/evidence/ |
run-experiment | Execute experiments against evaluation fixtures, capture pass/fail results. | .claude/skills/skill-scientist/scripts/run_experiment.py | $OUTDIR/results_manifest.json | $OUTDIR/evidence/ |
review-experiment | Score experiment results against success criteria, compute pass/fail threshold. | .claude/skills/skill-scientist/scripts/review_experiment.py | $OUTDIR/review_report.json | $OUTDIR/evidence/ |
promote-skill-update | Generate promotion proposals for passed experiments, describe required diffs. | .claude/skills/skill-scientist/scripts/promote_skill_update.py | $OUTDIR/promotion_report.json | $OUTDIR/evidence/ |
Evidence Required
- Save all generated JSON artifacts in the output directory.
- Include timestamps, workflow name, and tool versions in every output manifest.
- Store experiment logs (stdout/stderr captures) in the evidence directory.
- Evidence templates from
skill.yaml:hypothesis_generated:hypotheses.jsonwith gap IDs and priority levelsexperiment_result:results_manifest.jsonwith positive/negative control outcomesreview_passed:review_report.jsonwith normalized score >= 7.0promotion_ready:promotion_report.jsonwith target skill and diff description
References
- Source of truth:
skill.yaml - Runbooks:
runbooks/ - Evaluation harness:
.claude/skills/evaluation-harness/ - Inspired by: SakanaAI AI Scientist pattern (arXiv:2408.06292)
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

