deep-dive（深潜）

Deep Dive

Autonomous deep research using the same DAG-based planning pattern as Google's Deep Research — but running entirely on Claude Code with no external dependencies.

How it works

1. Plan — decompose the question into a DAG of sub-questions with dependencies
2. Fan out — run independent sub-questions in parallel via Agent subagents
3. Gap analysis — each subagent returns findings + identified gaps
4. Iterate — gaps become new sub-questions, fed back into the DAG
5. Synthesize — once all nodes complete, produce a final report

Steps

1. Decompose into a DAG

Given the research question, generate a DAG of sub-questions. Each node has:

• id: short identifier (e.g., q1, q2a)
• question: the specific sub-question to research
• depends_on: list of node IDs whose answers are needed first (empty = no dependencies)

Rules for decomposition:

• Start with foundational/context-setting questions that have no dependencies
• Build toward analytical/comparative questions that depend on foundational answers
• Aim for 4-8 nodes. If the topic needs more, cap at 12.
• Each node should be answerable with 1-3 web searches
• Questions should be specific enough that a researcher with no other context can answer them

Print the DAG as a table so the first brain can see the plan, then immediately proceed to execution — do not wait for confirmation.

Create a task for each DAG node using TaskCreate (description: the sub-question, status: pending). Also create tasks for "Gap analysis" and "Synthesize report". Update each task to in_progress when its wave launches and completed when the subagent returns. This gives the first brain real-time visibility into progress.

| ID | Question | Depends on |
|----|----------|------------|
| q1 | ...      | —          |
| q2 | ...      | —          |
| q3 | ...      | q1         |
| q4 | ...      | q1, q2     |

2. Execute in dependency order

Process the DAG in topological order:

Wave 1: Mark all Wave 1 node tasks as in_progress. Launch all nodes with no dependencies as parallel Agent subagents. As each subagent returns, mark its task completed. Each subagent gets this prompt:

You are a focused researcher. Answer this ONE question using web search:

Question: [the sub-question]

Instructions:
- Use WebSearch to find current, authoritative information
- Use 1-3 searches maximum
- Be specific and cite what you find

Return your answer in this exact format:

## Findings
[Your answer with specific facts, dates, numbers. Cite sources inline.]

## Gaps
[List anything you couldn't fully answer, contradictions you found, or
follow-up questions that would strengthen the answer. If none, say "None."]

## Sources
[List each source as: Title — URL]

Citation persistence: After each wave completes, append all sources from that wave to a file at /tmp/deep-dive-sources-[topic-slug].json as an array of {"node_id", "title", "url"} objects. This survives context compaction — if subagent results get compressed out of context, the sources file remains the source of truth. Read this file during synthesis to build the final Sources section.

Wave 2+: Once Wave 1 completes, mark all Wave 2+ node tasks as in_progress and launch nodes whose dependencies are now satisfied. Mark each task completed as its subagent returns. Include the findings from dependency nodes in the subagent prompt:

You are a focused researcher. Answer this ONE question using web search:

Question: [the sub-question]

Context from prior research:
[Paste findings from dependency nodes]

[same instructions as above]

Continue until all nodes complete.

3. Gap iteration (max 1 round)

Mark the "Gap analysis" task as in_progress. After all nodes complete, review the collected gaps across all subagents:

• If gaps are minor or don't affect the final answer: skip, move to synthesis
• If any gap is significant enough to change the conclusion: create 1-3 new targeted sub-questions and run them as a final parallel wave

Only do ONE gap iteration round. This is not an infinite loop.

Mark the "Gap analysis" task as completed when done (whether gaps were found or skipped).

4. Synthesize

Mark the "Synthesize report" task as in_progress. Combine all findings into a final report. Mark it completed when the report file is written. Structure:

## Deep Dive: [Topic]

### Executive Summary
[3-5 sentences: the key takeaway]

### Findings

#### [Theme/Section 1]
[Synthesized findings from relevant nodes, not just copy-paste]

#### [Theme/Section 2]
[...]

### Open Questions
[Anything that couldn't be resolved — be honest about what's still unclear]

### Sources
[Deduplicated list of all sources from all subagents]

Rules

• Always show the DAG plan first. Print it, then immediately start researching — no confirmation needed.
• Parallel where possible. Independent questions should always run as concurrent subagents.
• One gap round max. Don't spiral into infinite research loops.
• Synthesize, don't concatenate. The final report should read as a coherent document, not a list of subagent outputs stapled together.
• Be honest about confidence. If the research didn't produce clear answers, say so. Don't fill gaps with speculation.
• Always persist the final report. After synthesis, save the report as a markdown file in the appropriate project's docs/deep-dive/ directory (create it if needed). Determine the project root from the current working directory or the context of the research request. Use a slugified topic name with date as the filename (e.g., 2026-04-02-jira-docs-from-microservices.md). Never write final reports only to /tmp — they must land in a durable location within the relevant project.