Align generation prompts with review rubric to cut the ~70% repair rate (font sizing, design excellence)

[MarkusNeusinger]

Observation

During the 2026-06-10 bulk regeneration (pp-basic + 7 old specs, 15 libraries each, model sonnet), the window 2026-06-09T22:00Z..2026-06-10T06:30Z shows:

• 112 impl-generate runs, 78 impl-repair runs → roughly 70% of implementations needed at least one repair round (pp-basic: 6 of 7 new libs).
• Zero technical failures — every implementation eventually merged with final scores 84–93. The repairs were driven almost entirely by review feedback, not broken code.

That means the repair loop is currently functioning as a systematic second generation pass: each round costs one extra Claude generate run + one extra review run and adds ~20–30 min latency per implementation. At catalog scale (327 specs × 15 libraries) this roughly doubles LLM usage for predictable, recurring feedback.

The recurring rejection reasons

1. Font sizes below style-guide targets (VQ-01) — the most consistent pattern. Example: pp-basic/makie attempt 1 (PR feat(makie): implement pp-basic #8528) was rejected at 85/100 with titlesize=20 ≈ 40px effective on the 2400×2400 canvas vs the review rubric's ~67px title / ~53px axis-label targets. The reviewer even prescribed the exact fix (titlesize=28, labels 20, ticks 16). The generation prompts apparently don't state these numeric targets, but the review rubric scores against them.
2. Design-excellence deductions (DE-01 sophistication, DE-03 storytelling) — "correct but not elevated": missing visual hierarchy, no emphasis/annotation guiding the eye, no confidence band / deviation encoding. Reviews repeatedly suggest the same class of fixes.
3. Occasionally missed spec-required features (SC-02) — e.g. scatter-connected-temporal/muix attempt 1 (PR feat(muix): implement scatter-connected-temporal #8545) rejected at 81/100 with two spec-required features missing; final score after repair was 90.

Suggested direction (to be discussed)

• Add the rubric's numeric pixel targets per canvas size (3200×1800 / 2400×2400) to prompts/default-style-guide.md — ideally including the per-library translation (e.g. Makie titlesize≈28, CSS-px libs ~22px in the 1600×900 mount, etc.), so generation and review use the same numbers.
• Port the review rubric's DE expectations into the generation prompts as a short design-excellence checklist (one deliberate visual-hierarchy/storytelling element per plot).
• Add an explicit self-check against the spec's required features to the generation prompt before the snippet is committed.
• Optionally: record the repair-reason category in metadata so the effect of prompt changes on the repair rate is measurable.

Reference data

• Run window: 2026-06-09T22:00Z – 2026-06-10T06:30Z (pp-basic, scatter-connected-temporal, line-yield-curve, acf-pacf, recurrence-basic, funnel-meta-analysis, line-load-duration, area-elevation-profile)
• Example PRs: feat(makie): implement pp-basic #8528 (makie pp-basic, 85 REJECTED → 87 merged), feat(muix): implement scatter-connected-temporal #8545 (muix scatter-connected-temporal, 81 REJECTED → 90 merged), feat(highcharts): implement pp-basic #8521/feat(muix): implement pp-basic #8522 (highcharts/muix pp-basic, each one repair round)
• Final muix scores across the run: 84–91; highcharts: 84–91 — the repairs work, they are just predictable.