deep-research-skill

Deep Research Skill

A reusable research skill for general-purpose agents.

一个面向通用代理的可复用研究型 skill，可用于 Codex、OpenClaw、Claude Code 等环境。

This skill helps an agent turn vague research requests into a reliable decision, recommendation, or implementation plan instead of a pile of links. It is designed to be portable across runtimes such as Codex, OpenClaw, Claude Code, and similar agent systems.

It is designed for:

这个 skill 的目标是把模糊的研究需求转成可靠的结论、建议或执行方案，而不是产出一堆链接或资料摘抄。

deep research
technical investigation
tool, vendor, model, or architecture selection
codebase investigation
implementation planning
high-stakes recommendations that need evidence, tradeoffs, risks, and next steps

Example wording in this README is shown in English and Chinese, including requests such as deep research, deep dive, 研究, 调研, 深度调研, 选型, and 对比评估.

What's in this repo

SKILL.md: the main skill definition and operating workflow
AGENTS_SNIPPET.md: a compact version you can adapt into an AGENTS.md, system prompt, or equivalent agent instructions file
references/search-playbook.md: guidance for planning searches and validating sources
references/evidence-log-template.md: a lightweight evidence log template
references/subagent-playbook.md: role-pass and sub-agent workflow guidance

What the skill does

This skill pushes the agent to:

reframe the user's request as a decision or deliverable
identify constraints, assumptions, and evaluation criteria
verify volatile facts with current sources
search for limitations, counterexamples, pricing, deprecations, and risks
compare realistic options instead of summarizing search results
produce a recommendation, validation plan, and concrete next step

Triggering

Use this skill when the user needs researched comparison, investigation, synthesis, or a recommendation rather than a one-shot answer.

Typical trigger wording:

English: deep research, compare these, investigate, evaluate, recommend, pros and cons
Chinese: 研究一下, 调研一下, 深度调研, 比较一下, 评估一下, 帮我选型, 对比评估, 尽调

Avoid using this skill for:

simple factual lookups
translation or rewriting without research
straightforward coding tasks with little uncertainty

Depth modes

The skill supports three research depths.

Light

Use for narrow, low-risk questions where only a few facts matter and a short answer is enough.

Standard

Default mode for most research tasks. Use when the user needs a recommendation, comparison, or plan backed by multiple facts and tradeoffs.

Deep

Use when the stakes are high, the user explicitly asks for deep research, or the answer must hold up under scrutiny.

Examples of wording that should push the agent to Deep:

deep research
thorough analysis
due diligence
深度调研
深入分析
全面分析
详细评估
尽调

Installation / Integration

This repo is runtime-agnostic. How you install it depends on the agent system you use.

General integration options:

copy SKILL.md into your runtime's skill or prompt library
adapt AGENTS_SNIPPET.md into your repo-level AGENTS.md, system prompt, or equivalent instruction file
keep the references/ files alongside the main skill so the agent can open them when needed

For Codex, one convenient setup is to place this folder in the skills directory so the main file ends up at:

$CODEX_HOME/skills/deep-research/SKILL.md

One simple way:

mkdir -p "$CODEX_HOME/skills"
cp -R deep-research-skill "$CODEX_HOME/skills/deep-research"

For other runtimes such as OpenClaw or Claude Code, copy the same logic into the runtime's preferred skill, prompt, or repo-instructions mechanism.

Example prompts

Research whether we should use Postgres + pgvector or a dedicated vector database for our next release.
Do a deep dive on OCR vendors for Chinese invoices and recommend one.
研究一下当前这套鉴权架构，帮我找出风险和改造路径。
帮我做个深度调研：对比 Anthropic、OpenAI 和 Gemini 作为内部 coding agent 的长期方案。

Design goals

Prefer primary sources over summaries and marketing pages.
Separate facts, assumptions, judgments, and recommendations.
Right-size the output instead of forcing every task into a long report.
Support both true sub-agents and sequential role-based passes.
Stay portable across different agent runtimes instead of assuming a single platform.
Provide clear examples in both English and Chinese.

README