2.5 KiB
2.5 KiB
Iterative Fix-Verify Loop for Presentation QA
Pattern for achieving zero-defect presentations through repeated visual QA cycles. Derived from a 17-slide deck QA session (5 cycles, 6 issues → 0 issues).
The Loop
┌─────────────────────────────────────────┐
│ 1. GENERATE → node main.js │
│ 2. RENDER → soffice → pdftoppm │
│ 3. BATCH QA → batch_vision_qa.py │
│ 4. TARGETED → vision_analyze (deep) │
│ 5. FIX → patch source files │
│ 6. REPEAT 2-5 until ALL CLEAN │
└─────────────────────────────────────────┘
Cycle Progression (Real Example)
| Cycle | Issues Found | Action | Remaining |
|---|---|---|---|
| 1 | 1 (batch) + 6 (targeted) | Fixed all 6 | 3 remaining |
| 2 | 3 (targeted) | Fixed all 3 | 1 remaining |
| 3 | 1 (targeted) | Fixed | 1 remaining |
| 4 | 1 (targeted) | Fixed | 0 |
| 5 | 0 (full batch) | Done | 0 ✅ |
Why Multiple Cycles Are Needed
Fixes introduce new problems:
- Moving a subtitle down to fix overlap may push it into content below
- Enlarging a badge circle may overlap neighboring elements
- Changing line color for visibility may reduce contrast with another element
- Shortening a label to prevent wrapping may lose meaning
Never declare success after the first fix-and-verify. Always re-render and re-QA.
Batch + Targeted Prompt Division
Batch Prompt (Tier 1 — Breadth)
Generic, checks all slides quickly (~25s for 17 slides). Uses expanded 6-category prompt: overlap, overflow, contrast, layout, content, symbols.
Targeted Prompts (Tier 2 — Depth)
Element-specific questions for slides with:
- Diagrams: "Are the connecting lines visible? What color?"
- Badges/icons: "Do the letters inside circles render correctly?"
- Tables: "Are all cells populated? Any placeholder text?"
- Examples: "Does the example value match the description?"
- Arrows/connectors: "Do arrows render properly between elements?"
Efficiency Tips
- Only re-render and re-QA slides that were changed (not the entire deck)
- But ALWAYS do a full batch scan at the end to confirm no regressions
- Batch QA via direct API calls is 10-20x faster than sequential vision_analyze calls
- Save QA results to JSON for comparison between cycles