autoresearch

Self-improving AI-driven workflow that continuously refines, reviews, validates and optimizes research tasks. Use for iterative ongoing research.

SKILL.md

GitHub

Autoresearch (self-improving research loop)

A research agent that critiques and improves its own work across iterations.

Loop

  1. Plan — define objective + current best answer; maintain a working document research-state.md with: objective, findings-so-far (with citations), confidence per finding, open questions, next hypotheses.
  2. Execute — research the top open question (see deep-research method standards).
  3. Critique — adversarial self-review: weakest claims? unsupported leaps? recency issues? alternative explanations?
  4. Refine — fix, re-source, downgrade confidence where needed; update state file.
  5. Repeat until: all questions answered at target confidence, N iterations reached, or new info stops changing conclusions (convergence).
  6. Report — final synthesis + iteration log (what changed and why).

Rules

  • Persist state between runs; never re-research settled findings unless a contradiction appears.
  • Confidence scale: confirmed (2+ primary sources) / likely / tentative / unknown.

Output

Updated research-state.md each cycle + final cited report + convergence summary.