How to use the research agent skill: sourced answers before you code

What the research skill makes the agent do

When the next step depends on something the agent does not yet know, this skill tells it to stop and investigate. Its stated goal is to replace assumptions with evidence and to say plainly what remains uncertain. The skill is explicit that research answers questions and does not implement anything.

The agent works through a fixed sequence: it states the question precisely enough to know when it is answered, searches the repo and its docs first, consults official documentation for external SDKs, frameworks or APIs, and checks more than one angle, such as by file, by symbol, by caller, or by history. It then synthesizes its findings, each tied to where it came from.

• Narrow lookup: the agent answers directly.
• Multiple independent questions: the agent investigates them in parallel.
• Unknown-size discovery: the agent keeps going until additional passes surface nothing new.

Install the research skill with a ZIP file

Download the ZIP for the skill from the directory page, then unzip it. The skill ships as a single file named SKILL.md. Where you put the unzipped folder depends on how you want to use it.

For all your projects, place the skill in your personal skills folder. For a single project, place it in that project's .claude/skills/ folder so it applies only there. You can also upload the ZIP in the Claude apps under Settings → Capabilities → Skills.

• Personal (all projects): unzip into ~/.claude/skills/
• Single project: unzip into 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. Describe the question you need answered and say that you want it checked against the code or official docs. The skill is meant for questions where a wrong assumption would cost you, such as how an SDK really behaves or how an unfamiliar part of a codebase works.

Ask for the output you need. The skill's output format includes the question, findings with sources, what remains unknown, and a recommended next step if one follows. Asking for a plan before committing to code fits this format well.

• "Before we change the auth middleware, research how sessions are currently created and cited where."
• "Look up in the official docs how this SDK handles retries, and tell me what you verified versus inferred."
• "Investigate whether anything else calls the parser we plan to rewrite. Don't change code."

Read the output: sources, verified versus inferred

The skill requires a source for each finding. For code, that means a file and line; for external material, the document consulted. It also requires the agent to separate what it verified from what it inferred, to report contradicting evidence instead of picking the tidier story, and to list what it could not verify.

Treat the verified items as the basis for decisions and the inferred items as hypotheses to test. If the agent lists an open question, that is a signal to check it yourself before committing to a plan.

What the skill needs and its safety notes

The skill lists no requirements. It does not run scripts. It does use the network, because it consults official documentation for external libraries or APIs. Expect the agent to reach out for that material when a question involves an outside tool.

Because the skill forbids implementing as a side effect of research, it is designed to leave your code untouched. Still, read the agent's output before acting on it, since findings are only as good as the sources the agent could reach.

• Requirements: none listed.
• Runs scripts: no.
• Uses network: yes, for official documentation of external SDKs, frameworks and APIs.

Tips for getting better research

Give a precise question with a clear end point. The skill itself begins by making the question specific enough to know when it is answered, so a vague prompt produces a vaguer answer.

Use it for the questions that come before a plan, not for the implementation itself. If you want code written after the research, ask for that as a separate request.

• State what would count as an answer.
• Name the library or version you care about so the agent checks the right documentation.
• Ask for the list of open questions when you are about to commit to a plan.
• Request contradicting evidence explicitly if you suspect the codebase is inconsistent.