deep-research

deep-research

Execute one round of deep research from a brief produced by /create-brief.

A research project has this layout:

{output_dir}/
├── brief.md              # written by /create-brief
├── research/             # one file per topic
│   └── {topic-slug}.md
└── synthesis/            # syntheses across the corpus
    ├── general.md        # the whole-corpus synthesis, rewritten each round
    └── {axis-slug}.md    # per-axis cuts, owned by /synthesize

Each invocation runs one research round and stages the next:

1. Run the next pending round — dispatch one parallel Task agent per missing topic, each writing to {output_dir}/research/{topic-slug}.md.
2. Fully rewrite {output_dir}/synthesis/general.md — incorporating every round executed so far. Tight executive synthesis, not a table of contents in the body.
3. Propose gap topics for the next round — one-question-at-a-time conversation, then append ## Round N+1 to the brief and stop. The user re-invokes to execute the new round.

Why this matters

The user's "better too much research than too little" workflow only works if (a) topic agents run in true parallel — a single message with N tool calls, since sequential calls across messages run serially and waste hours; (b) each topic agent does fresh external web research rather than relying on training data; (c) the synthesis is rewritten each round so it integrates everything rather than accreting fragments; and (d) the brief is append-only so it stays an audit trail of what was researched and when.

Each topic file produced by an agent has this shape:

• 1500–4000 word Markdown file.
• 4–8 H2 sections, structured around the topic's natural axes (no generic template).
• Dense inline citations to external sources, each tagged with its type ([T1], [T2], [T3]) per the brief's Source types section.
• A ## Sources section at the bottom with full citation metadata.

The general synthesis has this shape:

• Markdown file, aim for 1500–4000 words — longer is fine when the material warrants it.
• Five H2 sections in this order: Contents (links to brief and topic files), Executive summary, Cross-cutting themes, Tensions and trade-offs, Open questions.
• Executive prose, not a list of facts. Navigation lives in the Contents section so the body can stay synthetic.