itsjwill/deep-dive

Deep Dive

A deep dive is not a web search. Find the best sources, read them for real, attack the idea from the other side, check the citations, and come back with an opinion and a plan.

Engine: RECON (MCP server auto-research) plus notebooklm-mcp. Without them, run lite mode (Step 3). Setup: references/setup.md. All tools: references/tools.md.

Rule zero: act first

When a trigger phrase appears, the FIRST action is research. No text first. No clarifying questions. Pull the topic out of context. If the topic is the user's own code, bot or business, research external strategies and best practices for that domain, then apply them to their situation in the brief.

Step 1: Check the library and the tools

• If the research tool is not in your tool list, search for it once (e.g. ToolSearch +auto-research research). Still missing means the server is not connected: go to lite mode (Step 3) and say so in one line.
• If the topic sounds familiar, run list_research_notebooks(). A matching recent notebook gets queried (Step 4) instead of rebuilt.

Step 2: Pick settings and run RECON

Depth by trigger:

Trigger	num_sources	Notebook questions	Extra
"research X"	5	3
"deep dive X"	6	5
"ultra think X"	10 (max)	6-8	Parallel angle agents (Step 4b)
"think the impossible X"	10	6-8	Step 4b + suggest_research() for adjacent ideas

Sources by domain (papers are noise on fast-moving tech, gold on science):

Topic type	sources	Add in parallel
Tech, tools, AI, software, marketing tactics	["youtube"]	X search + web search + GitHub search (recency matters)
Science, health, psychology, economics	["youtube", "papers"]
Trading, investing, niche craft	["youtube", "podcasts"] with mode="expert"	X search
Business strategy, general learning	["youtube", "papers", "podcasts"]	X search for founder/operator takes

X search (free, no account): web search with allowed_domains: ["x.com"]. Run two queries: the topic, and " doesn't work / overrated / mistakes". Results are threads, X articles and posts from people building the thing, often weeks ahead of YouTube. Skip X for science and health (mostly noise there).

Mode: mode="expert" for trading, bots, strategies or any niche where practitioners beat influencers (inverts popularity scoring). mode="general" otherwise.

Topic string: 3-8 searchable words. "liquidation cascade trading strategy", not a paragraph.

research(topic="<topic>", sources=[...], num_sources=N, mode="<mode>")

Takes 1-4 minutes: searches, scores each source for substance, credibility and transcript depth, swaps the weakest slot for a contrarian source, builds a NotebookLM notebook, updates the knowledge graph. Read status, notebook_url, notebook_id, sources, connected_notebooks.

Step 3: Handle failure, or run lite mode

Result	Do this
success / partial	Step 4 (note failed sources)
Login / "signed out" / "New notebook" error	Run notebooklm_login(). If signed out, tell the user in ONE line to run notebooklm_login(interactive=true), then lite mode. Never use notebooklm-mcp setup_auth (does not detect notebook.google.com)
No sources	Retry once with a broader topic string, then lite mode

Lite mode (no notebook): in parallel, run search_videos / search_papers if RECON is connected, plus 3-5 web searches (and the 2 X searches when the topic routes to X): the topic, " best practices / how practitioners do it", " problems / criticism / doesn't work", and " data / benchmark / results". Fetch and read the top 5-8 pages or transcripts directly. Keep at least one contrarian source. Then Steps 5-7 exactly as normal.

Step 4: Interrogate

4a. Notebook questions

Use notebooklm-mcp ask_question(question, notebook_url=<url>), one after another in the same session. Pull substance, not summaries:

1. Mechanism: "What is the core mechanism that makes this work? Step by step, with specifics."
2. Practitioner playbook: "What do the people getting real results do differently? Concrete tactics, numbers, settings, thresholds."
3. Contrarian: "Where does this fail? Biggest risks, criticisms and mistakes in the sources."
4. Numbers: "Every concrete number, benchmark, cost or result in the sources, each with the source that said it and a short exact quote."
5. Application: "Given <user's situation in one sentence>, what should they do first and what should they avoid?"

Deeper triggers add: "What do the sources disagree on, and who has better evidence?", "What non-obvious insight do most people miss?", "What would a 10x version look like?"

Vague answer = one follow-up that names the exact thing you need.

4b. Parallel angle agents (ultra think / think the impossible)

Spawn 3 subagents at once, each with a narrow brief and its own searches:

• Practitioner: what people with real results do, with numbers. Owns the X searches.
• Evidence: studies, benchmarks, data, official docs.
• Contrarian: failures, criticism, who says this does not work and why.

Each returns claims as claim | source URL | exact supporting quote. You merge. Do not let subagents write the brief.

Step 5: Verify before you write

Citations make weak claims look checked. Research shows deep-research tools fabricate 3-13% of URLs and only 40-80% of citations actually support the claim. So:

1. Pick the 5 claims the verdict depends on most.
2. For each, confirm the source says it: fetch the page, or ask the notebook "Quote the exact passage that says ".
3. Supported = keep. Not found = cut it or mark it (unverified). Contradicted = flip it into the contrarian section.
4. Only cite URLs that a tool returned in this session. Never write a URL from memory.

Step 6: Deliver the brief

Opinion first. A verdict, not a literature review.

## Deep Dive: <topic>

**Bottom line:** <2-3 sentences: what is true, what to do, confidence high/medium/low and why>

### What actually works
- <finding> [source]

### The numbers
| Metric | Value | Source |

### Where it breaks (contrarian view)
- <risk / failure mode> [source]

### Where sources disagree
<who says what, which side has better evidence> (skip if none)

### What this means for you
<applied to the user's real situation: first move, what to skip, what to test>

### Sources
| # | Title | Type | Score | Why it made the cut |
(mark the contrarian source; mark verified claims' sources)

Notebook: <notebook_url>  |  Verified: <n>/5 key claims

Flag hype with no evidence. If the evidence is thin, say the confidence is low. Do not pad.

Step 7: Offer next moves (one line each, only what fits)

• Keep querying the notebook (it stays live).
• suggest_research() for knowledge gaps; knowledge_map() for connections to past dives.
• After the user acts on it: track_edge(notebook_id, "edge" | "no_edge" | "negative_edge", notes). RECON learns which sources pay off; edge_report() shows the ROI.

Gotchas

• Empty key_concepts on every video = transcript extraction is failing (RECON < 1.0.1 with youtube-transcript-api >= 1.0). Scores then rank by title and channel only. Update RECON.
• NotebookLM moved to notebook.google.com. Logins lapse after ~12 idle days and Google may ask "Verify it's you". Fix with notebooklm_login(interactive=true).
• "Profile is open in another process" = a notebooklm-mcp query holds the Chrome profile. Wait a minute and retry.
• Papers off-topic (e.g. finance papers for a coding question): OpenAlex is weak on new tooling. Drop papers, use web + GitHub.
• More tool calls does not mean more accuracy. Fact-check accuracy drops as research sprawls. Five well-read sources beat thirty skimmed ones.
• X posts are leads, not proof. Engagement bait and confident wrong takes are common. Cite a post only when it shows real numbers or comes from someone who did the thing (builder, team member, operator); otherwise confirm it elsewhere.
• Reddit is unreachable for free. Reddit blocks Claude's web search, closed self-service API keys (Nov 2025) and returns 403 to scripts. Do not try to get around that. If the Claude in Chrome extension is connected and the user wants Reddit, open reddit.com search in a new tab and read the top 2-3 threads like a person would.
• One contrarian source is not balance. On debatable topics, actively state the strongest opposing case.