You MUST use a todo list to complete these steps in order. Never move on to one step if you haven't completed the previous step. If you have multiple read steps in a row, read them all at once (in parallel).

Add all steps to your todo list now and begin executing.

## Steps

1. [Read Deep Research Strategy Guide]: Read the documentation in: `./skills/sauna/[skill_id]/references/research.deep.guide.md` (Load topic decomposition strategy)

2. Clarify the research topic with the user. Ask:
1. What specific topic do you want to research?
2. What's driving this research? (decision, learning, preparation)
3. Preferred depth: Quick Overview (1-2 rounds), Moderate (3-4 rounds), or Comprehensive (5+ rounds)?

If user already specified the topic clearly, skip to decomposition.


3. INITIAL LANDSCAPE SCAN

Decompose the topic into 4-6 research branches using the guide. Common branch types:
- Fundamentals: Core concepts, definitions, how it works
- Players: Key companies, people, organizations involved
- Applications: Use cases, implementations, real-world examples
- Timeline: History, current state, future projections
- Controversies: Debates, risks, criticisms, open questions
- Economics: Market size, costs, business models

For each branch, formulate a specific search query.


4. PARALLEL SEARCH EXECUTION (Round 1)

Execute 4-6 web searches in parallel using @tool/websearch for each branch.
For each search:
- Use specific, well-formed queries (not vague terms)
- Extract 3-5 key facts with source URLs
- Note confidence level (High: multiple sources agree, Medium: single reliable source, Low: limited/conflicting info)


5. CREATE RESEARCH STATE

Write the JSON state to `./documents/tmp/research-map.json` with this structure:
```json
{
  "topic": "Topic Name",
  "depth": "Quick|Moderate|Comprehensive",
  "round": 1,
  "explored": [
    {
      "branch": "Fundamentals",
      "summary": "One-line key finding",
      "sources": ["url1", "url2"],
      "confidence": "high|medium|low"
    }
  ],
  "available": [
    {"branch": "Timeline", "preview": "What we'd learn if explored"}
  ],
  "suggested_deep_dives": [
    "Specific question raised by findings?"
  ],
  "history": [
    {"round": 1, "branches": ["Fundamentals", "Players"], "timestamp": "ISO date"}
  ]
}
```


6. CREATE RESEARCH MAP DISPLAY

Write the human-readable markdown to `./documents/tmp/research-map.md`:

```markdown
## Research Map: [Topic]

**Depth:** [Quick/Moderate/Comprehensive] | **Round:** 1 | **Sources:** [count]

### Explored
- [x] [Branch] — [One-line summary of key finding]

### Available to Explore
- [ ] [Branch] — [What we'd learn, suggested searches]

### Suggested Deep Dives
1. "[Specific question]" (based on [finding that raised this question])
2. "[Specific question]" (connects [branch A] + [branch B])

### Key Findings So Far
[2-3 paragraph synthesis of most important discoveries]

### Sources
- [Source title](URL) — used for [which claims]

---
**What should I explore next?** Pick branches above, suggest your own direction, or say "finalize" for the full report.
```

Present this map to the user and wait for their direction.


7. USER-DIRECTED EXPLORATION LOOP

Based on user selection:

**If user picks branches:** Use task tool to spawn `./skills/sauna/[skill_id]/references/recipes/research.deep.branch.md` in parallel for each branch. Provide: branch_name, parent_context, search_queries (2-4), depth_target. Subagents write to `./documents/tmp/branch-findings/*.json`. Wait for completion.

**If user asks a question:** Treat as new branch, spawn single subagent.

**If user says "continue":** Select top 2-3 suggested deep dives, spawn subagents.

**If user says "finalize":** Proceed to final report generation.


8. AGGREGATE FINDINGS

After subagents complete:
1. Read all files from `./documents/tmp/branch-findings/*.json`
2. Increment the round counter in `./documents/tmp/research-map.json`
3. Add newly explored branches to the "explored" array
4. Update "available" with remaining unexplored branches
5. Generate new "suggested_deep_dives" based on findings
6. Add this round to the "history" array
7. Update `./documents/tmp/research-map.md` with new state
8. Delete the processed files from `./documents/tmp/branch-findings/*.json`

Present updated map and repeat the exploration loop until user indicates satisfaction.


9. [Read Research Synthesis Guide]: Read the documentation in: `./skills/sauna/[skill_id]/references/research.deep.synthesis.md` (Load synthesis and report structure guidance)

10. GENERATE FINAL REPORT

When user is satisfied, synthesize all findings using the structure from the synthesis guide.
Save to `./documents/research-reports/*.md` with filename like {topic}-research-{date}.md

Include:
- Executive Summary (3-5 sentences)
- Key Findings by theme (not by search order)
- Timeline & Context (if explored)
- Key Players (if explored)
- Open Questions (honest about gaps)
- Confidence Assessment (claim-level and section-level)
- Sources (organized by topic, with descriptions)
- Reading time estimate (~250 words per minute)

Present the report location to user and offer to explore any remaining questions.


11. CLEANUP

Delete session files:
- `./documents/tmp/research-map.json`
- `./documents/tmp/research-map.md`
- Any remaining files in `./documents/tmp/branch-findings/*.json`