| name | search-topic |
| description | Do deep research on a topic and return a curated list of sources relevant to the book. Use whenever the user wants to find or research sources on a subject — "find papers on X", "what's the research on Y", "search for sources about Z", "any good sources on ...". Prioritises arXiv and other academic/primary sources but also returns Wikipedia (definitions/history), analyst and industry reports, and reputable blog/news articles. Returns a ranked list (topic, title, type, year, URL, why-relevant) — it does NOT download; pair with download-source next. |
| argument-hint | The topic or claim to research for the book. |
| user-invocable | true |
search-topic
Given a topic, find the sources worth pulling into the book and hand back a clean, actionable list.
This is the discovery stage — no downloading, no writing. The output is a shortlist a human (or the
research-topic orchestrator) can act on.
The book is AI-dō, a practical, evidence-led guide to working with AI; it favours primary sources
(it has deliberately replaced encyclopedia citations with primary ones) and values balance over
cheerleading. Search with that in mind.
What to return
A Markdown table, most-relevant first:
| Topic | Title | Type | Year | URL | Why relevant |
|---|
- Topic — the
sources/ topic folder it maps to (an existing one where it fits; otherwise propose a
new kebab-case topic name and say it's new). Existing topics include llm-foundations, ai-landscape,
agent-architecture, agent-security, spec-vs-vibe, ai-coding, code-quality, memory-context,
knowledge-work, personal-productivity, ambient-agents, agent-disciplines, governance-law,
human-ai-future, mastery-improvement, philosophy-ethics, software-engineering.
- Type —
arXiv, journal/conference, analyst report, wikipedia, blog/news, docs.
- URL — the canonical link: arXiv
abs page, DOI, or the publisher/original page. For arXiv, note the id.
- Why relevant — one line: the claim it supports or the balance it adds.
Below the table, add a short note on coverage and balance: which angle each cluster of sources takes,
whether you found a genuine counterweight, and any gap the search did not fill.
How to search
- Frame the topic against the book: which chapter/theme does it serve, and what claim is it meant to
support or challenge? A sharper frame gives sharper results.
- Search broadly, then rank by quality and fit. Priority order:
- arXiv and peer-reviewed / conference papers — primary evidence; prefer these. Capture the arXiv id.
- Analyst and industry reports (Stanford HAI AI Index, McKinsey, Deloitte, State-of-AI) — for
adoption/economics claims.
- Primary/official pages (lab blogs, standards bodies, product docs) — for what a lab/tool actually says.
- Wikipedia — only for definitions, history, or orientation, not as evidence for a contested claim.
- Reputable blog/news — practitioner writing, journalism; use with care and prefer the primary it cites.
- Verify the essentials at the source: exact title, author/org, year, canonical URL. Never rely on a
search snippet or a guessed author list.
- Seek a counterweight. If the topic is contested (most are), deliberately look for a source that
complicates or opposes the majority view — balance beats a one-sided pile.
- Filter hard. Drop SEO filler, content farms, and thin restatements. Prefer the primary over anything
that merely quotes it. Flag paywalled or hard-to-download items so download-source knows.
- Note likely overlap. If a candidate looks already covered in the book (the point is already cited),
say so — the reader may not need it.
Rules
- Never invent a title, author, year, or URL. If you cannot confirm it, fetch the page or leave it flagged.
- Prefer the published venue where one exists; fall back to the arXiv/preprint link.
- Relevance to this book is the bar, not general interest — a great paper off-topic for AI-dō is a no.