deep-research

这段内容是审核时保存的快照。外部来源才是完整且最新的版本。

---
name: deep-research
description: >-
  Routing front door for a structured, human-in-the-loop deep-research workflow
  (Weizhena/Deep-Research-skills) — turn a topic into an extensible outline, fan out parallel
  web-search agents to investigate each item into validated JSON, then render a complete markdown
  report. One skill, 4 reference pipelines: outline (research / add-items / add-fields), deep
  (parallel per-item investigation + field-coverage validation), report (TOC + per-field
  markdown), web-search (research agent + 5 routed source modules). Plugin: npx skills add
  https://github.com/akillness/jeo-skills --skill deep-research
allowed-tools: Read Write Edit Bash Grep Glob WebSearch WebFetch Task
compatibility: >
  Universal — every pipeline is plain Markdown instructions. The deep and
  report phases run two small Python helpers (validate_json.py needs PyYAML;
  the report generator is emitted per-run), so those phases need Python 3 +
  pyyaml. Web search uses each agent's native WebSearch (OpenCode requires
  OPENCODE_ENABLE_EXA=1). Works in Claude Code, Codex, Gemini CLI, OpenCode,
  Cursor, Copilot. Plugin-installable via `npx skills add`.
metadata:
  tags: deep-research, structured-research, research-outline, web-search, literature-review, benchmark, competitor-analysis, due-diligence, parallel-agents, report-generation, human-in-the-loop
  platforms: Claude, Codex, Gemini, Cursor, Copilot, OpenCode, All
  keyword: deep-research
  version: "1.0.3"
  upstream: https://github.com/Weizhena/Deep-Research-skills
  source: akillness/jeo-skills
  license: MIT
---

# deep-research

A routing-first front door for the **Deep Research** workflow from
[Weizhena/Deep-Research-skills](https://github.com/Weizhena/Deep-Research-skills)
— a two-phase, human-in-the-loop research method (outline generation, then deep
investigation) consolidated into one jeo-skill with **4 reference pipelines**.
Each pipeline ships its own prompt templates and output contract; this skill
classifies the request into the right phase, loads that pipeline, and executes
it exactly.

> One topic → one extensible outline → parallel per-item investigation into
> validated JSON → one complete markdown report. Every phase has a user
> checkpoint, so you keep precise control at each stage instead of handing the
> model a black-box "research X" prompt.

The reference pipelines hold the **dismantled-and-merged upstream skill text**
(per-command frontmatter stripped, headings nested, prompt templates kept
verbatim) — not a paraphrase — so each phase is self-contained here. Read the
matching pipeline reference before executing:
- [references/outline-pipeline.md](references/outline-pipeline.md) — `/research` · `/research-add-items` · `/research-add-fields` (Phase 1: generate + extend the outline)
- [references/deep-pipeline.md](references/deep-pipeline.md) — `/research-deep` (Phase 2: parallel per-item investigation + `validate_json.py` coverage gate)
- [references/report-pipeline.md](references/report-pipeline.md) — `/research-report` (Phase 3: TOC + per-field markdown report)
- [references/web-search-pipeline.md](references/web-search-pipeline.md) — the web-search agent + 5 routed source modules (github-debug · general-web · academic-papers · chinese-tech · stackoverflow)

## Plugin Installation
```bash
# This routing skill via jeo-skills (verified path)
npx skills add https://github.com/akillness/jeo-skills --skill deep-research

# Global install for one or more agents
npx skills add -g https://github.com/akillness/jeo-skills --skill deep-research -a claude-code -a codex -y

# Scripted install with knobs (Python dep + upstream slash-command skills)
WITH_DEPS=1 AGENTS="claude-code,codex" bash .agent-skills/deep-research/scripts/install.sh
```


The deep phase calls `scripts/validate_json.py` (needs `pip install pyyaml`).
The upstream repo also ships ready-made slash commands for Claude Code, OpenCode,
and Codex — `WITH_UPSTREAM=1 bash scripts/install.sh` clo