How to use the Deep Research agent skill (deep-research)

What the deep-research skill makes the agent do

The skill is a set of instructions that an agent follows when you ask for a thorough investigation. The agent first checks whether your question needs a full investigation at all. If one or two searches can answer it, the skill says to stop and just search directly.

For open-ended questions, the agent plans three to eight independent research angles, and parallel sub-agents each research one angle. The agent then checks the collected findings for gaps and, within a set budget, runs follow-up rounds. Finally, one agent writes a single report with inline citations, a list of open questions and a numbered source list, saved under a research folder.

• Research runs in parallel; writing is done by one agent in one pass.
• Depth modes: quick, standard (default) and deep, which set how many sub-agents and follow-up rounds are allowed.
• State is kept on disk in research/<slug>/ with brief.md, findings/ and REPORT.md, so work can resume after a interruption.

Install the deep-research skill from the ZIP

Download the ZIP for the skill from this page and unzip it. The skill folder contains SKILL.md and a reference folder with report.md, sources.md and subagent-prompt.md.

For a personal install, unzip it into ~/.claude/skills/. To make it available only inside one project, unzip it into that project's .claude/skills/ folder.

In the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

• Personal install: ~/.claude/skills/
• Project install: the project's .claude/skills/ folder
• Claude apps: Settings → Capabilities → Skills, then upload

Trigger the skill with a natural request

You do not need a special command. The skill is meant to activate when you ask for a thorough, multi-source investigation with a cited report. The source lists example phrases that match, in Chinese and English.

Use wording that signals depth, such as a request for a full study or a report. Words like quick or exhaustive can set the depth mode, according to the skill's instructions.

• "deep research on solid-state battery supply chains"
• "帮我全面研究一下" (help me research this thoroughly)
• "写一份调研报告" (write a research report)
• "多方求证" (cross-check from multiple sources)
• Add "quick" or "exhaustive" to change the depth mode

Examples of good and bad requests

The skill fits questions that need many sources and a decision-ready briefing. It does not fit quick facts. The skill itself says a single search is enough for simple lookups, and it directs academic literature surveys to a different auto-research skill.

• Good fit: comparing many options across several fields, such as "compare 20 frameworks" (the skill uses table-style batches for this).
• Good fit: a sourced briefing before a business decision.
• Poor fit: "What year was this library released?" (one search is enough).
• Poor fit: an academic literature survey (use the auto-research skill).

What the skill needs to run

The skill depends on a search tool, either the websearch tool or a search MCP server, plus WebFetch for reading pages. It also needs Bash so the agent can run the date command at the start, because the skill says never to assume the current year from training data.

The skill lists free external APIs in reference/sources.md, which do not need a key. The source file does not verify their specifics here, so check that file yourself before relying on any particular endpoint. If the search tool is unavailable, the reference describes a DuckDuckGo HTML fallback via WebFetch.

• A search tool: websearch or a search MCP server
• WebFetch
• Bash access to run date +%Y-%m-%d
• Free endpoints listed in reference/sources.md

Safety notes: scripts and network access

This skill runs scripts and uses the network. It sends search queries and fetches web pages, and it runs a Bash command to read the date. Review the files in the skill folder before installing if you need to control what runs on your machine.

The report relies on web search results, which can be wrong or out of date. The skill tells the agent to mark single-source claims and speculation, but a person should still check important claims against primary sources before acting on them.

• Runs scripts: yes. Uses network: yes.
• Writes files only under ./research/<slug>/ according to its workspace rules.
• Treat the report as a starting point that needs human verification for high-stakes decisions.

Tips for better reports

State your audience, time frame, region and the decision at stake in your request. The skill may ask one round of clarifying questions if the request is ambiguous, or it will write its assumptions into brief.md. Reading brief.md first lets you catch wrong assumptions early.

Check the Open questions and the [single source] and [speculative] flags in the report. These show where the evidence is thin.

• Give context up front to avoid guessed assumptions.
• Choose quick mode for a lighter pass and deep mode for contested topics; deep mode may ask you to confirm the angle list first.
• Read the executive summary, then follow the citations to the original pages.