CLAUDE.md, skill, hook, MCP or subagent: pick the right one

The steps

One step at a time. When a step’s check matches, tick it and the next one opens.

1. 01

   Step 1 of 3: See what each one loads, and when

   Context cost is the quickest way to tell them apart.

   Each feature uses Claude’s context in its own way. That cost is the first clue to which one fits.

   Loaded at session start

   • CLAUDE.md: the full text, on every request
   • Skills: their descriptions
   • MCP servers: tool names

   Loaded only when used

   • Skills: the full content
   • Subagents: a fresh context of their own
   • Hooks: nothing, unless they return context

   Check it worked

   You can say which one costs context on every request (CLAUDE.md) and which usually costs none (hooks).

   It worked — next step Not yet
2. 02

   Step 2 of 3: Ask five questions in order

   The first yes picks the feature.

   This is our rule of thumb, built from the docs’ triggers. Plugins sit outside it: they package the others for another repository.

   In order:

   1. You

      Must it happen every time?

      Yes: a hook

   2. You

      Should Claude always know it?

      Yes: CLAUDE.md

   3. You

      Is it a task you repeat?

      Yes: a skill

   4. You

      Does it need outside data?

      Yes: an MCP server

   5. You

      Would it flood the conversation?

      Yes: a subagent

   Check it worked

   Try it on one need from your own week. You land on one feature and can name the question that picked it.

   It worked — next step Not yet Back
3. 03

   Step 3 of 3: Sort six needs

   Pick a feature for each, then check the answers.

   1. Claude keeps running npm in a project that uses pnpm.
   2. You paste the same release checklist every Friday.
   3. The linter must run after every file Claude edits.
   4. Claude needs the open issues from your tracker.
   5. You want every caller of a function across hundreds of files, summarised.
   6. A second repository needs the same skills and hooks.

   your-project/

   • CLAUDE.md
   • .mcp.json
   • folder .claude
     • settings.jsonHooks go here
     • folder skills
       • folder release-checklist
         • SKILL.md
     • folder agents
       • find-callers.md

   Check it worked

   1. CLAUDE.md: a convention it got wrong.
   2. A skill: a playbook you repeat.
   3. A hook: it must run every time.
   4. An MCP server: data outside the repo.
   5. A subagent: lots of reading, short answer.
   6. A plugin: the same setup in another repository.

   It worked — finish Not yet Back