菜单
Orchestrates two-round parallel agent research with Exa orientation, breadth-first discovery, and depth-focused synthesis. Use when conducting comprehensive research, exploring complex topics requiring multi-source validation, or synthesizing findings from distributed investigation streams.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | deep-research |
| user-invocable | false |
| description | Orchestrates two-round parallel agent research with Exa orientation, breadth-first discovery, and depth-focused synthesis. Use when conducting comprehensive research, exploring complex topics requiring multi-source validation, or synthesizing findings from distributed investigation streams. |
Dictum: Iterative dispatch with inter-round critique maximizes research coverage.
Conduct comprehensive topic research via parallel agent dispatch.
Workflow:
exa-tools skill, map landscape, extract facetsparallel-dispatch skillDependencies:
exa-tools — Web search and code context queries (primary)perplexity-tools — Deep research and reasoning queries (supplementary)tavily-tools — Web search with domain governance (supplementary)parallel-dispatch — Agent orchestration mechanicsInput:
Topic: Domain to researchOutputPath: Target file path (passed by invoking command, default: report.md)Constraints: Context, scaffold, style from invoking skill[CRITICAL]:
OutputPath only—no other files.Dictum: Initial queries map landscape before dispatch.
Main agent executes exactly 3 Exa searches via exa-tools skill before dispatch.
Map domain landscape; identify facets for agent assignment. Produce facet list (6-10 independent research areas) for Round 1.
[IMPORTANT]:
exa-tools skill before dispatch.deep search tier for highest-quality results when topic warrants it.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: 6).
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]:
Dictum: Main agent builds skeleton—retains quality, identifies gaps.
Main agent (NOT sub-agent) processes Round 1 outputs.
| [INDEX] | [ACTION] | [CRITERIA] |
|---|---|---|
| [1] | Remove | Lacks focus, duplicates, no sources, pre-2025, fails quality |
| [2] | Retain | Addresses topic, sources, dates 2025-2026, converges across range |
Build skeleton from retained findings: [Domain N]: [findings] + Gaps: + Depth-Targets:
[CRITICAL] Skeleton is first corpus—Round 2 fleshes it out.
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:
| [INDEX] | [TYPE] | [PURPOSE] | [COUNT] |
|---|---|---|---|
| [1] | Focused | Specific gaps from skeleton | 4-6 |
| [2] | Wide | Broader context for areas | 2-4 |
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]:
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:
## [1][FINDINGS] — Synthesized research by domain## [2][CONFIDENCE] — High (convergent), Medium (single-source), Low (gaps)## [3][SOURCES] — All sources with attributionDictum: Gates prevent incomplete synthesis.
[VERIFY]:
exa-tools skillOutputPath by main agent onlyUse for Bash 5.2+/5.3 scripts, shell CLIs, entrypoints, CI jobs, cron, text-processing pipelines, ShellCheck remediation, and shell reviews. Enforces strict mode, dispatch-table routing, fork-minimal primitives, atomic I/O, signal-safe cleanup, and pragmatic functional shell patterns.
Enforces C# + LanguageExt functional/ROP style, type discipline, runtime-record effects, Scrutor composition, and module organization standards. Use when writing, editing, reviewing, refactoring, or debugging .cs modules, implementing domain models, sealed DU hierarchies, Eff/Fin/Validation pipelines, or configuring .csproj, analyzers, or DI.
Use for PostgreSQL 18 SQL, migrations, DDL, functions, RLS policies, indexes, query plans, extensions, and SQL embedded in TypeScript via @effect/sql-pg. Enforces set-algebraic queries, schema-level invariants, current PostgreSQL features, extension-first design, tenant security, observability, and migration safety.
Enforces Python + expression style, type discipline, error handling, concurrency, and module organization standards. Use when writing, editing, reviewing, refactoring, or debugging .py/.pyi modules, implementing domain models, ROP pipelines, Protocol-driven services, or configuring pyproject.toml, Ruff, or ty.
Enforces TypeScript + Effect functional/ROP style, type discipline, polymorphic surfaces, and module organization standards. Use when writing, editing, reviewing, refactoring, or debugging .ts/.tsx modules, implementing domain models, Effect services, persistence adapters, or boundary handlers, or configuring TypeScript, Effect, or lint/type-check posture.
Generates and validates project documentation: READMEs, ADRs, changelogs, ARCHITECTURE.md, and code documentation. Use when creating, updating, or reviewing any non-code markdown artifact, README file, architecture decision record, ARCHITECTURE document, CHANGELOG entry, or code documentation (docstrings, XML docs, TSDoc). Activates for: (1) scaffolding README, CHANGELOG, ADR, CONTRIBUTING, ARCHITECTURE, or SECURITY files; (2) writing or reviewing doc comments on exported APIs; (3) auditing documentation coverage or staleness; (4) generating release notes from commit history.