conducting-deep-research

# Conducting Deep Research

This skill uses an iterative methodology treating research as a diffusion process: start with noise (rough draft), apply guidance (research brief), denoise through cycles of critique → research → refine until quality converges.

Use TodoWrite to track these mandatory steps:

<required>
1. Clarify scope (if ambiguous)
2. Write research brief (guidance signal)
3. Generate initial draft from knowledge (noisy starting point)
4. Red team critique (identify noise)
5. Targeted web_search with reflection
6. Score sources and track contradictions
7. Refine draft (denoise)
8. Evaluate quality (score < 7? repeat 4-7, max 3 cycles)
9. Finalize
</required>

### Step 1: Clarify Scope

Ask me clarifying questions only if:
- Topic has multiple interpretations
- Key constraints missing (timeframe, geography, industry)
- Success criteria unclear

Skip for well-defined requests.

### Step 2: Research Brief (Guidance Signal)

```
RESEARCH BRIEF
Topic: [Specific question]
Scope: [Included/excluded]
Key Questions:
1. [Primary]
2. [Secondary]
3. [Tertiary]
Constraints: [Limits]
Language: [Match user's language]
```

### Step 3: Initial Draft (Noisy Starting Point)

Write from existing knowledge only, marking gaps:
- `[NEEDS VERIFICATION]` - uncertain claims
- `[RESEARCH NEEDED]` - missing information
- `[CONFIDENCE: LOW/MED/HIGH]` - flag uncertainty levels

This draft intentionally contains "noise" - the refinement loop will denoise it.

### Step 4: Red Team Critique

Attack the draft for:
- **Logic**: Circular reasoning, false dichotomies, unsupported claims
- **Gaps**: Missing perspectives, incomplete coverage
- **Sources**: Unsourced claims, single-source reliance

Rate each issue 1-10 severity. See [references/critique-framework.md](references/critique-framework.md).

### Step 5: Targeted Research with Reflection

**Search budget by complexity:**
- Simple verification: 1-2 searches
- Moderate topics: 3-5 searches  
- Complex research: 5-10 searches

**CRITI