Menu
Stateful time-bounded improvement loop with evaluator contract and dual logging
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Agent Communication Protocol — fan-out dispatch reference for Claude plugin orchestration
Disciplined 6-phase bug diagnosis loop. Build feedback loop, reproduce, hypothesise, instrument, fix with regression test, cleanup with post-mortem. Use for all bugs — bypasses PRD entirely. Ends with advisor-gate completion gate.
Interview-based planning skill. Ask one question at a time with recommended answers, cross-reference codebase, detect and defer to project-level planning conventions, and synthesize a concise plan that feeds vertical-slice. Use for features, small changes, and bug scoping.
Run the groundwork acceptance test harness. Documents how to test the plugin locally.
MANDATORY user acceptance testing before advisor-gate. Enforces real end-user testing — TUI via tmux/expect, API via real HTTP calls in docker-compose, web apps via Playwright browser. Unit/integration tests are INSUFFICIENT unless they exercise the system exactly as end-users experience it. No exceptions.
Engage maximum parallel fan-out mode. Use when you want to work 10x faster by dispatching all independent work simultaneously to specialist agents. Triggers on "ultrawork", "ulw", "fan out hard", "go parallel".
| name | autoresearch |
| description | Stateful time-bounded improvement loop with evaluator contract and dual logging |
| argument-hint | [--mission-dir <path>] [--max-runtime <duration>] [--resume <run-id>] |
Autoresearch is a stateful skill for bounded, evaluator-driven iterative improvement. It runs one mission at a time, iterates through non-passing results, and records each evaluation as durable artifacts. It stops only when the max-runtime ceiling or another explicit terminal condition is reached.
deep-interview --autoresearch to generate mission.md and evaluator.json — autoresearch does NOT generate its own evaluatorpass: boolean and optional score: number.pi/autoresearch/<mission-slug>/
mission.md ← what are we trying to improve or prove
evaluator.json ← evaluator command/script reference
runs/<run-id>/
evaluations/
iteration-0001.json ← {"pass": bool, "score"?: number, "notes"?: string}
iteration-0002.json
decision-log.md ← human-readable narrative per iteration
Each run gets a new <run-id> directory. Cron-scheduled reruns append new run dirs — never overwrite.
Verify artifacts: Confirm mission.md and evaluator.json exist under .pi/autoresearch/<mission-slug>/. If not, stop and instruct the user to run deep-interview --autoresearch first.
Record state: Create runs/<run-id>/ directory. Record: mission slug, iteration count (start: 0), started timestamp, max-runtime deadline.
Iterate — repeat until stop condition:
a. Run one experiment or change cycle based on mission.md guidance
b. Run the evaluator and capture its JSON output
c. Persist evaluations/iteration-NNNN.json (machine-readable)
d. Append a human-readable entry to decision-log.md (what was tried, what the evaluator returned, what to try next)
e. Continue unconditionally — pass: false is data, not a stop signal
Stop: When max-runtime is reached or user explicitly cancels. Do not self-terminate on pass.
Summarize: Read decision-log.md and produce a summary: how many iterations ran, trend in score if present, best result seen, recommended next steps.
The evaluator must output valid JSON to stdout:
{"pass": false, "score": 0.73, "notes": "routing accuracy improved but still missing security cases"}
Only pass is required. score and notes are optional but recommended for trend tracking.
## Iteration 0001 — <timestamp>
**Tried**: <what change or experiment was run>
**Evaluator result**: pass=false, score=0.73
**Observation**: <what the result tells us>
**Next**: <what to try in the next iteration>