Multi-Pass Review — Preventing Attention Dilution

Core 6 min +25 XP

💡

THE ANALOGY

Medical specialists vs one generalist reviewing 20 lab results. The cardiologist reviews cardiac results with full attention. The neurologist reviews neurological results with full attention. A generalist reviewing all 20 simultaneously misses the interaction between result 7 and result 15.

⚠️ EXAM TRAP — The Wrong Answer People Choose

Sending all changed files to one Claude call. With 15 files in one context, each file gets roughly 1/15th of effective attention — issues that per-file review catches get missed in the noise.

KEY POINTS

1 Attention dilution: reviewing many files at once reduces effective attention per file.

2 Multi-pass: per-file reviews (parallel, full attention per file) → cross-file synthesis (sequential).

3 Cross-file pass receives structured per-file summaries — NOT raw file contents.

4 Contradictions between per-file findings surfaced and routed to human resolution.

5 Two passes minimum — per-file correctness + cross-file integration.

Attention Dilution in Numbers

15 files in one call:
  Context: ~150k tokens total
  Effective per file: ~10k tokens
  File 13 competes with context from files 1-12

15 files, per-file (parallel):
  Context per call: full window
  Effective per file: FULL attention
  File 13: complete focus, no dilution

Multi-Pass Implementation

async def multi_pass_review(files: list[str]) -> dict:

    # Pass 1: per-file, parallel
    per_file = await asyncio.gather(*[
        call_claude(messages=[{
            "role": "user",
            "content": f"""Review {f} thoroughly. Focus ONLY on this file.

{read_file(f)}

Return JSON: {{
  "file": "{f}",
  "issues": [{{severity, description, line, fix}}],
  "interfaces": ["exported functions with signatures"],
  "assumptions": ["what this file assumes about external code"]
}}"""
        }])
        for f in files
    ])

    # Pass 2: cross-file synthesis, sequential, summaries not raw files
    cross_file = await call_claude(messages=[{
        "role": "user",
        "content": f"""Find cross-file integration issues ONLY.
Per-file reviewers already checked individual correctness.
Focus on: interface mismatches, contradictory assumptions, shared state bugs.

Per-file summaries:
{json.dumps(per_file)}

Return: {{
  "contradictions": [
    {{"function": "...", "file_a": "...", "file_a_says": "...",
      "file_b": "...", "file_b_says": "...", "human_question": "..."}}
  ],
  "integration_issues": [...]
}}"""
    }])

    return {
        "per_file": per_file,
        "integration": cross_file,
        "needs_human_resolution": extract_contradictions(cross_file)
    }

Key Takeaways

Attention dilution is real — per-file catches more issues
Parallel per-file, sequential cross-file — the canonical pattern
Cross-file receives summaries, not raw files — prevents re-dilution
Contradictions surfaced — route to human for resolution
Two passes minimum — individual + integration