True independence means zero generation context. Not “minimal” context, not “relevant” context — zero. Even sharing design goals from the generation prompt reduces findings by 67%.
The Isolation Spectrum
| Context shared with reviewer | Findings per file |
|---|---|
| Full generation conversation | 0.3 |
| Code + generation design goals | 1.2 |
| Code + review criteria only | 3.6 |
| Human reviewer | 4.0 |
Each leaked piece of generation context suppresses findings. “Design goals” seem harmless, but they tell the reviewer why the code was written that way — creating sympathy for the implementation rather than critical evaluation.
Full isolation (3.6) reaches 90% of the human baseline. Partial isolation (1.2) reaches only 30%.
What “Independent” Actually Means
The reviewer receives:
- The generated code
- Review criteria (from CLAUDE.md or the project’s quality standards)
- Test specifications
- Original requirements as a SEPARATE document (not the generation prompt)
The reviewer does NOT receive:
- The generation prompt
- Conversation history from the generation session
- Design rationale or architectural decisions from the generation phase
- Any reasoning about alternatives considered and rejected
Providing Requirements Without Generation Bias
If the reviewer needs to understand what the code should do, provide the original requirements document — the same document the generator worked from. Do not provide the generation prompt, which contains both requirements AND implementation reasoning.
The requirements document describes WHAT the code should accomplish. The generation prompt describes HOW Claude decided to accomplish it. The reviewer needs the former; the latter introduces bias.
One-liner: Provide the reviewer with the code, review criteria, and original requirements only — never the generation prompt or conversation history — because even “relevant” generation context suppresses findings from 3.6 to 1.2 per file.