Research

# User-invoked prompt · research (v1.8.1)

> Multi-agent parallel research → synthesized wiki draft with built-in
> counter-bias check + decoupled CitationAgent verification. ROUTER reads
> this when the user wants to seed a new wiki entry from external sources.

## Trigger keywords

- `调研一下 X` / `深挖 X` / `多角度研究 X`
- `research X` / `deep dive X`
- `/research <topic>` (slash command escape hatch with arguments)

## Argument parsing

ROUTER extracts:
- `topic` (required, free-form string)
- `depth` ∈ {`normal`, `deep`} (default `normal` = 5 agents; `deep` = 8)
- `thesis` (optional string; if provided, agents weigh for/against)

If natural language ambiguity (e.g. "research stablecoins"), ROUTER may
ask 1 clarifying question (which angle, what scope) before launching.
Limit clarification to ONE round to keep total time ≤ 7 minutes.

## Pre-flight

1. Verify cwd is a second-brain repo (`meta/` + `wiki/` exist). If not,
   refuse and tell user `/research` writes wiki entries; must run inside
   a second-brain vault.
2. Read `wiki/SCHEMA.md` to confirm frontmatter contract.
3. List `wiki/` domain directories. Cache list — synthesis must place the
   draft under one of these (or explicitly create new with user approval).
4. Estimate cost upfront: tell user "Spawning N agents — estimated 6-9
   min and ~$0.40-1.00 in API tokens (includes citation verifier).
   Continue?" Wait for "yes" or "go". (Skip estimate prompt if user
   explicitly typed `/research`; that's already opt-in.)

## Phase 1 · Decompose (main context, ~30 sec)

Decompose `<topic>` into N angles where N=5 (normal) or N=8 (deep):

**Normal (5 angles)**:
1. **academic** — peer-reviewed papers, formal definitions, established models
2. **practitioner** — industry blog posts, conference talks, current implementations
3. **contrarian** — known critiques, failed implementations, opposing schools
4. **origin** — historical first-principles, who coined the concept and why
5. **adjacent** — neighboring fields that touch the topic from outside

**Deep adds (3 more)**:
6. **mechanistic** — how does it actually work at the mechanism level (math, data flow, protocol)
7. **data-statistics** — quantitative evidence, benchmark results, market sizes
8. **meta-review** — what surveys / systematic reviews already exist

For each angle, write a 1-sentence brief (e.g. "academic: find peer-reviewed
papers from 2020-2026 on X; identify the 2-3 most-cited"). Cache angles for
phase 2 + show user the angle list before launching agents (1 line each).

## Phase 2 · Parallel research (5-8 subagents, ~5 min hard cap)

In a SINGLE message, launch all N subagents in parallel (one Task tool call
per agent within the same message — Claude Code parallelism requirement).

Subagent type: `general-purpose`

Each subagent receives: