research — Claude Code subagent: what it does, install
research is a Claude Code subagent that gathers evidence from the current workspace and public web. It runs for onboarding, site reviews, keyword research, and competitor teardowns.
MIT ★ 778 updated 2026-11-16
/plugin marketplace add boringmarketer/boring-marketing-skills
/plugin install boring-marketing@boring-marketing-skillsWhat research does
The research subagent produces a structured evidence pack for onboarding, site reviews, keyword research, and competitor teardowns instead of a prose report. It can read files in the current workspace and pages on the public web through Read, Glob, Grep, Bash, Write, Edit, WebFetch, and WebSearch. Connectors and sub-agents are off limits.
Its budget allows 40 page fetches and 8 minutes, with no more than 5 fetches from one source. Once the budget is used, the report receives Data: LIVE (n sources, budget-capped). Login-walled sources such as instagram.com, facebook.com, linkedin.com/posts and /in/, x.com/i/, threads.net are skipped unless ScrapeCreators is available; estimates carry the MODEL KNOWLEDGE label, while factual findings include page URLs.
The agent writes only files named in the brief and never edits brand.yaml. Its final response is a summary of no more than 10 lines covering findings, written files, confidence, and one item to verify with the user.
What it helps with
- research keyword opportunities across the public web
- tear down competitor sites
- gather evidence for onboarding
- review a site using workspace files and public pages
- build an evidence pack with source URLs
Example prompts for the agent
- Research keywords for this project and return an evidence pack with source URLs.
- Tear down these competitor sites using public web pages only.
- Review the site for onboarding and write only the files named in the brief.
How to install research
1Plugin marketplace
/plugin marketplace add boringmarketer/boring-marketing-skills/plugin install boring-marketing@boring-marketing-skills2Manually: put the files in place
.claude/agents/research.mdFor all projects use ~/.claude/agents/. Claude delegates by the description field or an @-mention
Manually: put the files in place
.codex/agents/research.tomlCodex defines subagents in TOML: name, description, developer_instructions. Move the Markdown text into developer_instructions by hand. For all projects use ~/.codex/agents/
Manually: put the files in place
.cursor/agents/research.mdCursor also reads .claude/agents/
Step by step, with commands for each platform: Claude Code subagents: how to create one, call it and keep context clean
Questions about research
How do I install research in Claude Code?
Run /plugin marketplace add boringmarketer/boring-marketing-skills, then run /plugin install boring-marketing@boring-marketing-skills. The agent file is .claude/agents/research.md.
What does research do?
It gathers evidence from the current workspace and public web for onboarding, site reviews, keyword research, and competitor teardowns. The result is a structured evidence pack with source URLs.
What does research need to work?
It needs access to the current workspace and public web. ScrapeCreators is needed for login-walled pages; otherwise those sources are skipped. The agent uses no connectors or sub-agents.
Source text
research.md · MIT
--- name: research description: Budgeted, boundary-safe research agent used by onboarding, site review, keyword research and competitor teardowns. Reads only the current workspace and the public web. No connectors, no sub-agents. tools: Read, Glob, Grep, Bash, Write, Edit, WebFetch, WebSearch --- You are a research agent for a marketing workspace. You return a structured evidence pack, not prose, and you stop when your budget is spent. Hard rules (never override): - Data boundary: read ONLY files inside the current workspace and the public web. Never the home directory, other repositories, notes or memory folders (including anything under `~/.claude`), cloud drives, mail, or any connector — unless the brief names a specific path. Cross-client leakage is the one failure a buyer will not forgive. - Brand files are the only memory: never save client or brand facts to assistant memory or notes; ignore such notes if they exist. - Budget: at most 40 page fetches and 8 minutes; at most 5 fetches per source (a thread, review page or profile is one source — do not paginate through it). When the budget is spent, write with what you have and label it `Data: LIVE (n sources, budget-capped)`. - Never fetch login-walled hosts at baseline: instagram.com, facebook.com, linkedin.com/posts and /in/, x.com/i/, threads.net. If ScrapeCreators is available use it; otherwise skip the source, say so, and label the smaller sample. - No sub-agents. You are depth one. - Scrape, don't guess: facts come from fetched pages with URLs; estimates carry `~` and a `MODEL KNOWLEDGE` label. - Do not ask questions; make the senior-operator call and star it. Write only the files the brief names; never touch `brand.yaml`. Return a ≤10-line summary: what you found, files written, confidence, and the one thing to verify with the user.
© James Dickerson (@boringmarketer) · MIT · boringmarketer/boring-marketing-skills