| id | autoresearch |
| name | Autoresearch |
| version | 2.0.0 |
| layer | skill |
| runtime | nodejs |
| author | home23 |
| description | Iteratively improve a weak skill through score, tweak, and retest loops. Run against a target skill, define failure mode and rubric, execute the loop, and get a scored report with recommendations. |
| category | meta |
| keywords | ["skill","improve","autoresearch","optimize","iterate","quality"] |
| triggers | ["improve this skill","autoresearch this skill","optimize the skill","why is this skill weak"] |
| capabilities | [{"autoresearch_loop":"Run a full score/tweak/retest loop against a target skill"}] |
| requiresTools | ["skills_run","skills_list","skills_get","skills_audit"] |
| dependsOn | ["skill-loader"] |
| composes | ["knowledge-structuring","source-validation"] |
Autoresearch
Use this skill when a skill exists but performs inconsistently and needs deliberate improvement rather than one-off fixes.
When to use
Use autoresearch for:
- a skill that triggers inconsistently across similar asks
- a skill whose instructions feel vague, brittle, or underspecified
- quality work where the target is the skill itself, not just the current answer
Workflow
- Define the failure mode clearly.
- Choose a score rubric (5 dimensions, 1–5 each).
- Run representative prompts against the skill across N rounds.
- Auto-revise
SKILL.md after each round (targets weakest dimension).
- Stop when score gain flattens (≤ 0.2 improvement).
- Return a scored report with per-round results and recommendations.
Score Rubric (5 dimensions)
| Dimension | What it measures |
|---|
queryStrategy | Detects zero/low results and auto-broadens or retries |
qualityFiltering | Filters spam, elevates high-engagement signal |
resultCoverage | Gets meaningful result volume per query |
actionContract | Returns well-formed, predictable responses |
documentation | SKILL.md examples and gotchas are clear and actionable |
Scores: 1 (broken) → 3 (functional) → 5 (excellent).
autoresearch_loop action
{
"action": "autoresearch_loop",
"input": {
"targetSkill": "x-research",
"failureMode": "X search returns spam for health queries with no retry logic",
"promptSet": [
"sauna cold plunge contrast therapy",
"infrared sauna health benefits",
"cold plunge recovery sleep"
],
"scoreRubric": {
"queryStrategy": { "target": 4 },
"qualityFiltering": { "target": 3 },
"resultCoverage": { "target": 3 },
"actionContract": { "target": 4 },
"documentation": { "target": 3 }
},
"maxRounds": 3
}
}
Output
Returns a JSON report with:
rounds[]: per-round scores per prompt and dimension averages
summary: start/final/total gain, rounds run, stopped-early flag
recommendations[]: dimensions still scoring < 3 after loop
Report saved to workspace/reports/autoresearch/autoresearch-{skill}-{timestamp}.json.
Gotchas
- Do not autoresearch every miss. Only use it when the pattern repeats.
- Keep the prompt set stable while you iterate or the score signal becomes noise.
- Fix routing text and gotchas before inventing more actions.
- Autoresearch runs inside the target skill's context — it does NOT need its own auth.
- Round scores start at ~2–3 for a broken skill and should climb with each revision.