research for Claude Code: a skill that saves web research as sourced documents

research is a skill that investigates a topic on the web and files the result as a sourced document in one repository. It is a single skill with no public repository, so this post covers the problem it solves, who should build one, the layout to start from, and the setting you should keep when you do.

What problem it solves

Ask an agent to look something up and you usually get a good answer that disappears with the session. A week later you need the same numbers for a slide, and the sources are gone. research changes where the work ends. Every run produces a Markdown document with a one-line summary, key data, insights, and a source list, saved under a date-prefixed file name in a single research repository no matter which project you called it from.

It also controls cost by splitting research into two tiers. The default is lightweight research with three to five web searches. It moves to deep research, which runs a multi-agent harness, only when you explicitly ask for it. When a page blocks the default fetch, it tries a bypass wrapper and then browser automation before giving up, so a single blocked source does not end the run.

Who it fits

The chapter rates four situations. Three are enough to decide.

Situation	Fit	Reason
Building a documented library of work knowledge	Good fit	Fixed template and storage location
Gathering evidence for classes and presentations	Good fit	Records sources and access dates
A short question you will ask once and discard	Poor fit	Proceeds through saving a file

Verifying disputed numbers for a publication sits in the middle. The skill can do it, but that usually calls for the deep tier and its token bill.

Build it and run it

Because there is no public repository, you build the skill yourself. This is the structure the chapter gives, with the date-prefixed file name pattern written in English.

# Example folder layout
~/.claude/skills/research/
  SKILL.md                      # Full procedure
research-hub/                   # Repository for results
  templates/research-template.md
  researches/                   # YYYY-MM-DD-topic.md
  researches/assets/            # HTML and PNG diagrams for each document
  scripts/render-diagram.sh     # Convert HTML diagrams to PNG
  scripts/enqueue-wiki-ingest.sh
  scripts/research-to-notion.py

It activates when you call it by name or ask for something like “Research this” or “Find out about this topic.”

/research Vector database adoption in Korea in 2026

The run follows a fixed order. The skill first defines the scope and asks two or three narrowing questions if the topic is broad. It picks the lightweight tier unless you signaled deep research, and in that case it asks for approval before spending. It then writes the document from the template, turns the main insight into one HTML diagram rendered to PNG, and verifies the file before recording it.

What trips people up

The helper files live outside the skill folder. SKILL.md calls a template and scripts for diagram conversion, wiki queueing, and Notion recording, but the skill folder itself holds only SKILL.md. If you copy the skill alone, those steps fail. Either prepare the template and scripts alongside it, or delete the steps you do not need, and change the repository path and Notion settings to your own while you are in the file.

Where this comes from

This post adapts part of chapter 01-06 of The Claude Code & Codex Skills Guidebook. The full chapter covers the complete stage table, including the wiki check that cites existing pages before any new search, and the trade-offs of fixing every result to one hub repository, and the book has 52 more skills like this one. Read the full chapter in The Claude Code & Codex Skills Guidebook on Leanpub.