deep-research（深入研究）

[H1][DEEP-RESEARCH]

Dictum: *Iterative dispatch with inter-round critique maximizes research coverage.*

Conduct comprehensive topic research via parallel agent dispatch.

Workflow:

1. §ORIENT — Execute 3 Exa searches via exa-tools skill, map landscape, extract facets
2. §ROUND_1 — Dispatch 6-10 agents for breadth coverage via parallel-dispatch skill
3. §CRITIQUE_1 — Filter findings, retain quality, build skeleton with gaps
4. §ROUND_2 — Dispatch 6-10 agents to flesh out skeleton
5. §CRITIQUE_2 — Synthesize holistically, deduplicate, produce final output

Dependencies:

• exa-tools — Web search and code context queries
• parallel-dispatch — Agent orchestration mechanics

Input:

• Topic: Domain to research
• OutputPath: Target file path (passed by invoking command, default: report.md)
• Constraints: Context, scaffold, style from invoking skill

[CRITICAL]:

• [ALWAYS] Main agent writes to OutputPath only—no other files.
• [ALWAYS] Sub-agents RETURN structured text—main agent is sole file writer.
• [NEVER] Sub-agents use Write, Edit, Bash, or create any files.

[1][ORIENT]

Dictum: *Initial queries map landscape before dispatch.*

Main agent executes exactly 3 Exa searches via exa-tools skill; these map topic structure.

Map domain landscape; identify facets for agent assignment. Produce facet list (6-10 independent research areas) for Round 1.

[IMPORTANT]:

• [ALWAYS] Execute 3 Exa searches via exa-tools skill before dispatch.
• [ALWAYS] Extract facet boundaries from results.
• [NEVER] Dispatch before orient completes.

[2][ROUND_1]

Dictum: *Breadth via parallel dispatch—6-10 agents exploring independent facets.*

Dispatch 6-10 sub-agents via parallel-dispatch. Assign each agent unique scope from orient facets.

Agent Count: Scale by task complexity (default: 8).

Agent Prompt:

Scope: [Specific facet from orient]
Objective: Research this facet comprehensively
Output: Return structured text (CRITICAL → FINDINGS → SOURCES)
Context: [Topic background, constraints]
Constraint: DO NOT write files—return text only

[CRITICAL]:

• [ALWAYS] Dispatch ALL agents in ONE message block.
• [ALWAYS] Include "DO NOT write files" constraint in every agent prompt.
• [NEVER] Create overlapping scopes.

[3][CRITIQUE_1]

Dictum: *Main agent builds skeleton—retains quality, identifies gaps.*

Main agent (NOT sub-agent) processes Round 1 outputs.

Skeleton: Build from retained → [Domain N]: [findings] + Gaps: + Depth-Targets:

[CRITICAL] Skeleton is first corpus—Round 2 fleshes it out.

[4][ROUND_2]

Dictum: *Depth via parallel dispatch—same agent count, focused on skeleton gaps.*

Dispatch 6-10 sub-agents (same count as Round 1) via parallel-dispatch.

Agent Assignment:

Agent Prompt:

Scope: [Gap or depth-target from skeleton]
Objective: [Focused: fill gap | Wide: broaden context]
Output: Return structured text (CRITICAL → FINDINGS → SOURCES)
Context: [Skeleton content—build on, don't repeat]
Prior: [Relevant Round 1 findings]
Constraint: DO NOT write files—return text only

[CRITICAL]:

• [ALWAYS] Same agent count as Round 1.
• [ALWAYS] Include skeleton context.
• [ALWAYS] Include "DO NOT write files" constraint in every agent prompt.

[5][CRITIQUE_2]

Dictum: *Main agent synthesizes holistically—final corpus for downstream use.*

Main agent (NOT sub-agent) compiles final research output and writes to OutputPath.

Integrate: Merge Round 2 → skeleton. Cross-reference rounds. Resolve conflicts (prioritize sourced, current, convergent).

Filter: Remove duplicates, out-of-scope content, superseded items, unresolved conflicts.

Write: Single file to OutputPath with structure:

• ## [1][FINDINGS] — Synthesized research by domain
• ## [2][CONFIDENCE] — High (convergent) | Medium (single-source) | Low (gaps)
• ## [3][SOURCES] — All sources with attribution

[6][VALIDATION]

Dictum: *Gates prevent incomplete synthesis.*

[VERIFY]:

• Orient: 3 Exa searches executed via exa-tools skill
• Round 1: 6-10 agents in ONE message, all returned text (no file writes)
• Critique 1: Skeleton built, gaps identified
• Round 2: Same count, focused on skeleton, all returned text (no file writes)
• Critique 2: Final synthesis, duplicates removed
• Single file written to OutputPath by main agent only