Convergence speed is the most underrated quality signal in AI content tooling. Not output quality on round one, not the richness of the rubric, not how many dimensions the scorer checks. How fast does the system land on something good and stop? If your AI content tool is still "improving" a draft after five rounds, one of two things is true: either the rubric has no ground truth, or the refinement prompt is rewarding itself for iterating. Neither is useful to the person trying to post something this afternoon.
We hit this problem ourselves. Habitu Studio is the internal multi-tenant social content tool we built to manage content for Habitu's own marketing and for the restaurant brands we dogfood with. The voice-check feature, which runs a Claude pass over each draft to score it against a stored brand-voice rubric and surface refinement suggestions, was looping 7 times or more per draft before settling. It wasn't broken. The output was genuinely better than the input by round 7. But "7 rounds of Claude" is not a UX; it's a billing event and a trust problem.
What Was Causing the Loop
The original voice-check prompt returned a single blob of feedback: praise for what was working, findings for what wasn't, and a list of suggested edits interleaved with affirmations. The refinement step that followed took the full blob as input and rewrote the draft against all of it.
The problem is that mixed praise and findings create an unstable optimization target. When you tell a model "here's what's good, here's what's bad, now improve it," the model has no clean signal for what to hold and what to change. It frequently drifted: fixing a real issue while inadvertently softening a phrase that had already been scored as a strength. Round 3 would then flag the drift, and the loop would restart. We were essentially asking a model to satisfy two contradictory signals simultaneously and expressing surprise when it oscillated.
The second cause was that the Apply action wasn't selective. The UI had a single "Apply suggestions" button that fired against the full findings list, including findings that were soft stylistic preferences rather than clear voice violations. Applying a preference-level note would shift the draft slightly, which would trigger a new check, which would find something new to flag. Infinite refinement, zero convergence.
The Fix: Three Changes
Separate praise from findings at the prompt level. The scorer now returns two discrete fields: a strengths array (hold these) and a findings array (change these). The refinement prompt receives only the findings. It is explicitly instructed not to touch phrases or structures that appear in the strengths list. This alone cut average iteration rounds from 7+ down to around 4.
Classify findings by actionability before surfacing them. Each finding now carries an actionable boolean. A finding is actionable if it points to a specific, reversible change: a prohibited term, a structural violation, a missing message pillar. A finding is non-actionable if it's a degree-of-tone note ("could be slightly more direct") that a model can't operationalize without guessing. Only actionable findings get passed to the refinement step.
Apply only fires on actionable items. The UI change matched the prompt change. The Apply button no longer fires against the full findings list. It fires against the actionable subset only. If the only remaining findings are non-actionable, the button renders as disabled, which is the correct UX signal: you're done, the model just has opinions now.
The third fix also required a useCallback dependency correction. The Apply handler had a stale closure over the initial findings array, which meant that after round 2, clicking Apply was replaying round 1's findings against a round 2 draft. That's a subtle bug that looked like a quality issue. It was a wiring issue. The commit message was "voice check: fix useCallback deps + Apply fires on actionable items only", both fixes in one pass.
What Convergence Looks Like Now
The system now runs a check, returns separated strengths and actionable findings, applies the actionable set, runs one confirmation check, and terminates if actionable findings are zero. In practice that's 2 rounds for drafts with 1-2 clear voice violations, 3 rounds for drafts that were more substantially off-rubric to start. Round count scales with the gap between draft and rubric, which is the correct behavior. A well-written on-brand draft exits after a single confirmation pass with no changes.
We also tightened the iteration budget as a hard cap. The prompts now include an explicit instruction: if fewer than 2 actionable findings remain after the current round, do not recurse. This backstop costs nothing on well-behaved drafts and catches edge cases where the actionable classifier misfires on an ambiguous finding.
Why This Matters Beyond Studio
The pattern generalizes. Any AI content tool that uses iterative refinement against a scoring rubric will exhibit the same failure mode if the rubric mixes signal types and the apply action isn't selective. The instinct when quality is inconsistent is to improve the rubric — add more dimensions, be more specific, expand the examples. That instinct is usually wrong. The fix is almost always upstream: separate what the model should hold from what it should change, and give the apply action a clear stopping condition.
AI content tools that require 6 rounds to converge aren't "more thorough." They've externalized their stopping criterion onto the user. The user runs the tool until it feels done, which is not a product.
What We Shipped
The voice-check improvements landed as part of Studio v2, which shipped at v0.37.0.0 on 2026-05-03 (PR #152). The relevant commit trail: "voice check: separate praise from findings," "Apply only fires on actionable items," "voice check: fix useCallback deps." The iteration count drop was immediate and measurable, we went from flagging 7-round sessions in the cost log to 2-3 rounds as the new ceiling across the drafts we've run since.
The cost side matters too. Seven Claude API calls per draft at production volume is not a rounding error. Three calls is. The quality improvement was real, but the cost improvement is what makes the feature viable at scale for a multi-tenant tool serving multiple brand voices simultaneously. Correctness and economics pointed in the same direction, which is usually a sign you fixed the right thing.
Studio is the internal tool we use to create content for Habitu's branded restaurant channels. The voice-check feature is part of the content generation layer that ensures every post we publish from a restaurant brand stays on-rubric without a human editor in every loop. Getting the loop to converge fast was a prerequisite for the feature being useful in production at all.