| name | research-deck |
| description | Research a list of topics and produce a sourced slide deck with per-slide speaker notes. Use when asked to build a lecture/deck from current AI-industry topics. Encodes the whole orchestrator-worker + QA pipeline. |
| argument-hint | ["deck title"] |
| allowed-tools | Read, Write, Bash, Agent, WebSearch, WebFetch |
/research-deck — research → artifact, in one command
You are the orchestrator of a research-to-artifact loop. Turn topics.md
into a sourced deck (out/deck.json) plus rendered out/slides.md and
out/speaker-notes.md. Deck title: $ARGUMENTS (default "State of Agentic AI").
Pipeline
- Read
topics.md — one research topic per line.
- Research (parallel workers). For each topic, dispatch a
researcher
subagent (they run independently — this is the parallelization pattern). Each
writes out/research/<slug>.json and must satisfy
schema/topic-research.schema.json: ≥3 key points, every point carrying a
real source URL, plus a headline and takeaway. Verify each with
scripts/verify-research.sh; if a topic fails, re-run that worker (bounded).
- Assemble. Build
out/deck.json against schema/deck.schema.json: a title
slide, one content slide per topic (claims = the topic's sourced key points,
speaker notes from its headline + takeaway), and a summary slide. Then render
out/slides.md and out/speaker-notes.md. scripts/build-deck.sh does this
deterministically; use it, or do it yourself with editorial judgment.
- QA loop (evaluator-optimizer). Run the
deck-qa subagent (or
scripts/qa-deck.sh out/deck.json). It enforces the rules: every slide has
speaker notes, every content slide has ≥1 claim, every claim cites a source,
every topic is covered. If it reports problems, fix the deck and re-QA. Do
not finish until QA passes — the Stop hook will block you otherwise.
Rules
- An unsourced claim never reaches a slide. If research can't source it, drop it.
- Every slide gets speaker notes — no exceptions.
- More iterations cost more tokens (research is the expensive part). Don't loop
for its own sake; stop as soon as QA passes.
