How to use the Deep Research Orchestrator agent skill

What is an agent skill and what does this one do?

An agent skill is a folder of instructions (here, a file named SKILL.md) that an AI agent loads when a task matches the skill's description. It tells the agent how to work, not just what to answer.

This skill is a research workflow. It asks the agent to clarify the scope, plan sub-questions that can be researched independently, run those as parallel sub-agent sessions, check important claims adversarially, and synthesise the findings into a report. Each finding is expected to carry a citation to its source.

Install the deep-research skill

Download the ZIP for the deep-research skill from this page. Then choose one of the following options.

For Claude Code, unzip it into your user skills folder ~/.claude/skills/ for use everywhere, or into a project's .claude/skills/ folder to limit it to that project.

In the Claude apps, upload the ZIP under Settings → Capabilities → Skills.

• Personal use across projects: ~/.claude/skills/
• Single project: the project's .claude/skills/ folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Trigger the skill with a natural request

You do not need a special command. The skill is meant to load when a task needs multiple sources, traceable citations and fact-checking. Describe the goal in plain words and say what you expect back.

Stating the audience, time range, region or output language helps, because the agent may ask two or three clarifying questions before starting. If you have already settled those points, say so and the agent should not keep asking.

• Literature review: "Do a deep research review of published work on sodium-ion battery cathodes since 2020, with sources I can trace."
• Competitor survey: "Research the three main competitors to our product and produce a cited report comparing pricing models and stated features."
• Technology selection: "Compare vector database options for a 10 million document search system, with cited evidence and flagged gaps."

What the agent does during a run

The skill describes a fixed sequence. The agent first clarifies scope if the question is unclear, then splits the topic into sub-questions with clear output standards. Independent sub-questions are dispatched at the same time as sessions named for research and fact-checking roles.

Key claims, numbers and conflicting findings are sent to a separate fact-checking session, which marks each as supported, uncertain or refuted with sources and confidence. Claims that fail the check are left out or clearly downgraded. The agent then writes the synthesis itself rather than pasting sub-agent output together.

By default the full report is saved as an .html file in the project's working folder and presented as a deliverable. The chat reply only summarises it. If you ask for Markdown, plain text or JSON, the skill follows your format instead.

• Clarify scope only when the question is unclear
• Research sub-questions in parallel where they do not depend on each other
• Check key claims and flag conflicts
• Save the report as HTML by default and present it as a file

What the skill needs to run

The skill depends on specific tools and features in your environment. Without them the workflow cannot run as designed.

It uses a web_search tool, a present_artifacts tool and a write_file tool. It also needs sub-agent session support, meaning the create_session, wait_inputs and get_result operations. It does not run scripts, but it does use the network.

• web_search tool
• present_artifacts tool
• write_file tool
• Sub-agent session support (create_session, wait_inputs, get_result)
• Network access: yes
• Runs scripts: no

Safety notes: network use and verification

The skill sends web searches over the network and reads the results it finds. It runs no scripts itself. Its citations come from the model's own retrieval and checking, so the skill itself does not verify sources against an independent database.

Treat the report as a well-organised draft that points you to sources. For any decision that matters, open the cited sources and check the key numbers yourself.

• No scripts are executed by this skill
• Web content is retrieved over the network, so read sources critically
• Human checking of cited sources is still needed for high-stakes use

Tips for better research reports

Give the agent a precise question and the decision the report will support. Name the audience and the language you want. Ask for gaps to be flagged if you care more about honesty about missing evidence than a tidy conclusion.

Expect the run to be slow and to use many tokens. Use the skill for questions where the depth is worth the cost, and use a normal chat answer for quick facts.

• State the decision the report must support
• Specify the time range and region
• Ask explicitly for evidence gaps and uncertainty to be marked
• Keep quick factual questions out of this workflow