Research

Before asking Q7, load references/plugins/INDEX.md to get the current plugin list. Build Q7 from the index entries, listing each plugin with its user-facing description and prerequisites. Example rendering (adapt to actual index contents):

7. Would you like to activate any integrations for this session?

   Available:
   • NotebookLM — Google NotebookLM: indexes all sources and answers cross-source
     questions via RAG; produces a second briefing-doc output artifact.
     Requires: notebooklm-py>=0.6 installed; Google account.

   Reply with the name(s) of any you'd like to use, or "none" to skip.

If the index has no entries (empty table), omit Q7 entirely.

After the user answers Q7: For each plugin the user activated, load its file from references/plugins/ and run its Section 1 (Setup & Availability Check) before proceeding. If setup fails and the user chooses not to fix it, mark that plugin as unavailable and continue with the standard workflow. Plugins that pass setup are active for the entire session.

After intake, confirm your understanding:

"I'll research [restate question], focusing on [scope], using [source types]. [If any plugins active: I'll use [plugin names] — [one-line summary of each plugin's role].] I'll flag gaps and missed angles at the end. Ready to start?"

Step 2 — Source discovery

Identify relevant sources. Read references/source-types.md to guide tool selection and discovery strategy per resource type (web, local files, arXiv, code repos, video).

Aim for breadth first: 4–8 diverse primary sources before going deep on any one.

If NotebookLM is enabled: Follow references/plugins/notebooklm.md Section 2 instead. The NLM research agent (source add-research --mode deep) is the primary discovery mechanism. Use WebFetch-based discovery only as a fallback if fewer than 4 sources are returned.

Step 3 — Primary source analysis

Fetch and analyse each primary source. For each source:

1. Extract the key claims, data, and evidence relevant to the research question.
2. Apply the quality rubric (see references/quality-rubric.md) and assign a quality annotation.
3. Note any secondary leads the source points to (citations, linked resources, related repos).

Do not follow secondary leads yet — collect them for Step 5.

If NotebookLM is enabled: Follow references/plugins/notebooklm.md Section 3 instead. Use notebooklm source guide <id> to get NLM's AI summary and keywords per source as a substitute for direct WebFetch reads. Apply the quality rubric as normal based on that output. Also run notebooklm generate mind-map after all sources are loaded to get a concept overview.

Step 4 — Synthesis

Combine findings across primary sources:

• Group by theme or sub-question.
• Surface agreements, contradictions, and patterns.
• Note claims that are single-sourced (unverified) vs corroborated.
• Do not silently choose one side of a conflict — document it.

If NotebookLM is enabled: Follow references/plugins/notebooklm.md Section 4 instead. Issue the core synthesis queries via notebooklm ask, generate a data table for structured comparisons, and produce the NotebookLM briefing doc (Artifact A). Use the RAG responses with their inline citations as the primary evidence base for this section.

Step 5 — Secondary source check-in

Present the secondary leads identified in Step 3 and ask the user for approval before fetching:

I've identified these secondary sources worth following:

1. [Title / URL] — [why it's relevant, e.g. "cited by three primary sources on X"]
2. [Title / URL] — [why it's relevant]
3. ...

Shall I fetch and analyse these? Approve all, select a subset, or skip.

Wait for the user's response. Then fetch approved sources and integrate findings into the synthesis.

If NotebookLM is enabled: After user approval, add each approved URL via notebooklm source add <url> before fetching. Follow references/plugins/notebooklm.md Section 5 to re-query the expanded notebook.

Step 6 — Gap retrospective

After synthesis, evaluate the research against this checklist and flag every gap found:

• Opposing view — Is there a credible counter-argument or minority position?
• Recency — Are the most recent developments (last 6–12 months) represented?
• Practitioner vs theoretical — Is both academic/theoretical and real-world/practitioner perspective present?
• Geographic / cultural variation — Does the finding generalise, or is it region/culture-specific?
• Adjacent domains — Are there insights from neighbouring fields that apply here?
• Negative results — What has been tried and failed? Is that documented?
• Stakeholder perspectives — Are all relevant stakeholders (users, operators, critics) represented?

Write "none identified" only if genuinely true after checking all seven.

If NotebookLM is enabled: Before running the checklist, issue the seven gap-framing queries via notebooklm ask (see references/plugins/notebooklm.md Section 6). Use the answers as evidence when evaluating each checklist item. Then save the full session history to the