Skip to content

/neuroflow:autoresearch

Infinite improvement loop — point it at any file(s) and it runs a worker-evaluator cycle indefinitely, keeping or reverting each change based on whether it improved the artifact.

Inspired by Andrej Karpathy's autoresearch. Runs until you interrupt it.

When to use it

  • You want to improve a hypothesis, paper section, grant aim, or analysis plan overnight
  • You want to leave something running and come back to a better version
  • You want a live dashboard showing quality trends across iterations
  • You want to explore what "continuous improvement" looks like for a research artifact

How it works

First run — initialization

  1. Claude determines the active phase from project_config.md
  2. You name the files to improve (or use --target path/to/file.md)
  3. Criteria are built in three layers:
  4. Phase defaults (e.g. paper: language, claim support, statistics, reproducibility, novelty)
  5. Context-inferred (e.g. adds "aligns with preregistered hypotheses" if preregistration/ exists)
  6. Your additions (you can add any criteria before the loop starts)
  7. A baseline snapshot is saved to history/v000/
  8. A local dashboard server (server.py) is generated
Dashboard: python .neuroflow/{phase}/autoresearch/server.py → http://localhost:8765

The loop — never stops until you interrupt

Each iteration: 1. Worker — makes one focused change to the tracked files (targets the weakest criterion) 2. Evaluator — compares the new version to the previous best and returns BETTER / WORSE / NO CHANGE 3. Keep or revert — if BETTER, the new version is archived to history/vNNN/; if WORSE, the previous best is restored 4. Log — result logged to results.md with verdict, delta, and next focus

Invocation forms

Form Behaviour
/autoresearch Uses active phase from project_config.md
/autoresearch paper Targets the paper phase explicitly
/paper autoresearch Any phase command + autoresearch keyword triggers this
/paper autoresearch --target manuscript/intro.md Pre-fills the tracked file

Dashboard

python .neuroflow/{phase}/autoresearch/server.py
# → http://localhost:8765
# → http://localhost:8765?watch=1   (auto-refresh every 30s)

The dashboard shows: - Quality curve — running delta over iterations (staircase: rises on KEPT, flat on REVERTED) - Numeric metric charts — power, R², rejection rate, etc. when applicable - Last recommendation — what the evaluator says to target next - Plateau warning — triggered after 5 consecutive reversions

Files created

.neuroflow/{phase}/autoresearch/
├── flow.md
├── program.md         # task + criteria (edit this to guide the loop)
├── __thetask__.md     # pointer to tracked files
├── results.md         # iteration log
├── server.py          # dashboard server
└── history/
    ├── v000/          # baseline
    ├── v001/          # first KEPT improvement
    └── ...

Files read and written

Direction Files
Reads .neuroflow/project_config.md, .neuroflow/flow.md, tracked external files, program.md, results.md, history/ snapshots
Writes .neuroflow/{phase}/autoresearch/, tracked external files (on KEPT), history/vNNN/ snapshots
  • neuroflow:autoresearch skill — full protocol, criteria, and dashboard template
  • /paper — uses the worker-critic loop (bounded, 3 iterations) for section drafting
  • /pipeline — multi-step orchestration across phases