deep-dive
Skill1 starsChanged 7 months agoMITView on GitHub
Claude-native deep research using DAG-based query planning, parallel subagent execution, and gap-driven iteration. No external API needed.
25 other projects use a version of this.
Adopted by
Most recently taken up by
- TechNickAI/hermes-skillstheir own version, first seen Nov 16, 2026
- ZaxbyHub/opencode-swarmtheir own version, first seen Nov 15, 2026
- okou-ai/okou-skillstheir own version, first seen Nov 11, 2026
- hom3bru/skill-belttheir own version, first seen Nov 8, 2026
- buddyh/agent-skillstheir own version, first seen Nov 8, 2026
- myths-labs/musetheir own version, first seen Nov 6, 2026
Used with
Projects with it also use ralplan, deep-interview and 8 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
ibahgatSkillralplan
Create a strict, quality-first execution plan with explicit gates and rollback-ready sequencing.
Also in 4 projects that use this
klauswgSkilldeep-interview
在明确批准执行之前,通过数学化歧义门控进行苏格拉底式深度访谈
Also in 3 projects that use this
thoughtbotSkilltrace
Run a scenario and record what actually happened.
Also in 3 projects that use this
sahinSkillexecute
Also in 4 projects that use this
snarktankSkillralph
Convert PRDs to prd.json format for the Ralph autonomous agent system. Use when you have an existing PRD and need to convert it to Ralph's JSON format. Triggers on: convert this prd, turn this into ralph format, create prd.json from this, ralph json.
Also in 3 projects that use this
nicolasestremSkillbrainstorm
Structured idea generation and solution exploration with systematic analysis.
Also in 4 projects that use this
Other versions25 projects use a version of this.
Word for word · 1
- SkotAlexsander/skills★ 1100% same
Edited · 0
None found.
Wrote their own version · 24
- zereight/gitlab-mcp★ 2.1ksame name
- zaxbysauce/opencode-swarm★ 532same name
- ZaxbyHub/opencode-swarm★ 531same name
- Joonghyun-Lee-Frieren/oh-my-antigravity★ 214same name
- Joonghyun-Lee-Frieren/oh-my-gemini-cli★ 214same name
- okou-ai/okou-skills★ 81same name
- vm0-ai/vm0-skills★ 81same name
- myths-labs/muse★ 38same name
- joyehuang/skills★ 9same name
- bis-code/claude-toolkit★ 8same name
14 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository2 other files it gives its agents.
- news-from-discussionsSkill
- secret-guardSkill
The file
We split this skill into the rules it gives an agent
6 rules. 0 are common, 6 are distinctive, and 0 are very specific to this project.
0
Common
Most of the 23 edited versions we read keep these.
6
Distinctive
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 30
Start with foundational/context-setting questions that have no dependencies
- Distinctive: line 98
If gaps are minor or don't affect the final answer: skip, move to synthesis
- Distinctive: line 99
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
- Distinctive: line 101
Only do ONE gap iteration round.
- Distinctive: line 134
One gap round max: Don't spiral into infinite research loops.
- Distinctive: line 136
Be honest about confidence: If the research didn't produce clear answers, say so.
“Common” is measured: kept by at least 6 in 10 of the 23 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
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
- Plan — decompose the question into a DAG of sub-questions with dependencies
- Fan out — run independent sub-questions in parallel via Agent subagents
- Gap analysis — each subagent returns findings + identified gaps
- Iterate — gaps become new sub-questions, fed back into the DAG
- 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.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

