Investigate a specific question in official documentation, source code or specifications and save cited findings as Markdown.
Install and prepare
Run the following in your project terminal, then use the prompt below in Claude Code. This installs editable, project-scoped skill files.
npx skills@latest add mattpocock/skills --skill research --agent claude-code
For Codex, replace claude-code with codex in the command and /research with $research in the prompt. With the complete Claude Code plugin, select /mattpocock-skills:research instead; avoid installing the same skills twice.
You need access to documentation or source code and an agent environment that supports delegated work. Specify the decision, technology versions and destination directory so the research can feed directly into implementation.
Your first task
Adapt the scenario and paths to your project, then enter this in the agent conversation:
/research
Compare IndexedDB and OPFS for an offline notes app: transactions, file access and browser support. Use primary sources and save the findings to docs/research/offline-storage.md.
Work through the task
Turn a topic into questions
Browser storage alone is too broad. State the data shape, offline needs and target browsers: text plus attachments, multiple editing tabs and exportable backups, for example. Ask how each constraint affects the choice instead of requesting a general survey.
Tie conclusions to primary material
The workflow delegates reading and follows claims to their owning source, such as a specification or product documentation. Ask it to distinguish documented behavior from assumptions, especially for compatibility, quotas and experimental features. Search snippets alone are not enough for implementation decisions.
Use the note in the next decision
Review whether each question was answered and open the sources that drive the decision. Convert unresolved points into small experiments, such as saving and reading an attachment in the target browser. A researched recommendation does not mean that experiment has already been run in your app.
Check the result
The file should contain the question, findings, supporting links and open issues in the requested directory. Another developer should understand the decision basis without reconstructing the entire conversation.
Common questions
The report is generic. Add concrete constraints and options you need to rule out.
Does it monitor future changes? This is a single research task; recurring updates need a separate workflow.
No subagents are available. Ask the current agent to follow the same research standard; parallel reading depends on the host.
Turn research into the next decision
Frame the research around a constrained question rather than collecting links. For offline notes, specify the target browsers, multi-tab writes, attachment sizes, and backup needs. The report should compare options against those conditions and separate documented behavior from what must still be tried in your target environment. Each decision-changing claim should link to an original document, standard, or source file. The end should name a recommendation, open questions, and a small experiment that could settle them.
The Skill delegates reading to a background task. Confirm that exactly one task appears for the question; stop an accidental duplicate rather than accepting two overlapping reports. Ask for the absolute path to the Markdown result and open it. A spoken summary with no file, or two important claims leading only to secondary write-ups, is not yet a useful handoff to developers. Date short-lived facts such as compatibility or prices rather than treating them as permanent rules.
Scope the research deliverable
Decide on the deliverable before invoking the skill. For a single API parameter, a direct answer from the official docs may suffice. For substantial reading that another session must inspect, explicitly request a cited research note. Community feedback describes routine lookups triggering the heavier background-research workflow. State the decision question, required sources, output path, and what evidence would change the decision so a long report does not answer the wrong question.