The Fourth Draft Was the One. You Stopped at Two.
Self-Critique In-Thread Degrades
Ask an AI to critique its own draft in the same conversation and you hit a structural problem: the context that produced the draft also produces the objections. It finds the flaws it was already positioned to see, and each round tends to elaborate rather than improve.
The usual stopping rule is whoever gets bored first. That is not a quality bar.
Running the loop in an isolated session breaks the attachment to the draft. Measuring convergence replaces the boredom heuristic with a number you can read.
How It Actually Works
Every call advances exactly one cycle, so you can read the critique and stop early if a revision goes sideways:
- State the quality bar — not the topic, the standard the draft has to clear
- Draft — a first pass against that bar, in a session of its own
- Critique — domain-aware, and not written by the part that is attached to the draft
- Revise — one cycle, then report how much actually changed
- Converge — stop when the changes go cosmetic, not when the patience runs out
A converged draft is not automatically a good one. Convergence means the process stabilized — if it stabilized on something mediocre, the fix is a sharper quality bar, not more rounds.
When to Reach for This
- ✓A spec, ADR, or incident report that has to survive an adversarial read
- ✓External copy where the first draft is never the deliverable
- ✓You cannot tell whether another pass would help
- ✓In-thread self-critique has started elaborating instead of improving
Use Cases
Architecture Decision Records
Refine until a new engineer could act on it without asking a follow-up question.
Incident Write-Ups
Tighten the failure narrative before it reaches people who were not in the room.
Public-Facing Copy
Iterate past the version that merely sounds fine, and know which pass got you there.
Other Reasoning Tools
Different failures need different interventions. Pick the right tool for the moment.
Your AI Polishes Until You Stop Asking. This One Knows When to Stop.
$20/month for a refinement loop that reports convergence instead of guessing at it.