| name | autoresearch |
| description | Run a measured improvement loop for a skill by defining a yes/no checklist, capturing a baseline, scoring candidate outputs, and logging each iteration's keep/revert decision. Use when a skill works inconsistently and you want evidence-driven iteration instead of random prompt edits. |
| version | 1.0.0 |
| author | Vivi |
| tags | ["research","iteration","evaluation","prompt","skill-quality"] |
| platforms | ["all"] |
| dependencies | ["python3"] |
autoresearch
Turn the method into a stable workflow. This skill does not pretend to be a fully autonomous optimizer yet; instead it gives the team a reproducible workspace for iterative skill improvement with checklist scoring, baseline capture, and iteration logs.
Use this skill when
- A skill works 70% of the time and fails the other 30%
- You want measured iteration instead of ad hoc prompt editing
- The team needs a checklist-driven loop with changelog and baseline evidence
- A future agent should be able to resume from previous iteration history
Workflow
- Initialize an
autoresearch/ workspace.
- Define 3-6 yes/no checklist questions.
- Lock a test input and record the baseline.
- Make one change only.
- Score the new output.
- Keep or revert the change and append the result to the changelog.
Commands
Initialize a workspace:
python3 main.py init --root /tmp/autoresearch-demo --skill-name "landing-page-skill" --author Developer_Lead --questions "标题是否具体?|是否没有 buzzwords?|CTA 是否具体?"
Score one run:
python3 main.py score --skill-name "landing-page-skill" --answers y,n,y
Log one iteration:
python3 main.py log-iteration --root /tmp/autoresearch-demo --iteration iter-01 --change "Add explicit headline rule" --before 56 --after 72 --keep
Inputs and outputs
- Input: target skill name, checklist questions, yes/no answers, iteration changes, baseline notes
- Output:
autoresearch/ workspace with checklist.md, baseline.md, CHANGELOG.md, runbook.md, and optional JSONL scoring logs
Validation
Minimum validation commands:
python3 -m py_compile main.py
python3 main.py score --skill-name "landing-page-skill" --answers y,n,y
Constraints
- This is a research loop scaffold, not a fully autonomous optimizer
- Checklist quality determines result quality; vague questions will poison the loop
- Only change one thing at a time, or the changelog stops being useful
