- name
- autoreason-lite
- description
- Apply a bounded multi-candidate self-refinement loop (A/B/AB + judges + do-nothing option) to improve drafts, plans, and analyses while preventing scope creep. Use when user asks to improve writing/strategy/code explanations iteratively with quality control.
Autoreason Lite
Use this skill when a user asks to "improve this," "refine this," or "make this better" and quality matters more than one-shot speed.
What this does
A bounded refinement tournament:
- A (incumbent): current draft unchanged
- B (adversarial revision): deliberately different attempt
- AB (synthesis): merge best parts of A and B
- Judging pass: pick winner with explicit rubric
- Convergence rule: stop early if A keeps winning (no-change is valid)
This reduces over-editing and drift.
Default operating profile
- Max rounds: 3
- Judges: 3 independent judge personas
- Aggregation: Borda-like ranking (1st=2 pts, 2nd=1 pt, 3rd=0)
- Convergence: stop if A wins 2 rounds (or winner unchanged 2 rounds)
- Length guardrail: output within ±15% of requested length unless user asks otherwise
- Voice lock: preserve user's tone profile (technical / founder / viral) unless asked to shift
When to use
Use for:
- Long-form writing
- Strategy memos
- Explanations/tutorial drafts
- Product copy
- Decision frameworks
Avoid for:
- Deterministic factual extraction
- Tiny edits user already specified exactly
- Time-critical one-liners unless user requests deep refinement
Execution steps
- Clarify success criteria (tone, audience, length, goal) if missing.
- Generate candidate B from A:
- Must change structure or argument order (not just wording). - Must preserve critical facts.
- Generate candidate AB:
- Keep strongest parts from A and B. - Remove redundancy.
- Run 3 judges independently with rubric:
- Accuracy / faithfulness - Clarity - Usefulness for user goal - Concision / scope control
- Score candidates; choose winner.
- Repeat up to max rounds with winner as new A.
- Return final with short “what changed + why”.
Judge prompt template
Use references/judge-rubric.md.
Output format to user
- Final refined result
- 2-4 bullets: key improvements
- Optional: one-line note if loop stopped due to convergence (no meaningful gain)
Quick presets
- Technical: precise wording, fewer claims, concrete mechanisms
- Founder: outcomes + positioning + credibility signal
- Viral: short lines, strong hooks, high readability, no fluff
Safety + quality constraints
- Never invent facts to make prose sound better.
- Keep user intent stable unless explicitly asked to pivot.
- Prefer no-change over noisy edits.
- If confidence drops, surface uncertainty instead of bluffing.