Chapter 04 · Section II · 12 min read
Explore, Plan, and the others
Claude Code ships with a small family of specialised subagents. Each one is a persona with a specific tool set and a specific job. Choosing between them well takes about five minutes to learn and saves hours over the following months.
A subagent is not just “a fresh Claude.” Each built-in agent type comes with an opinionated tool set and system prompt tuned for a particular kind of work. Picking the right type is a design decision — like picking the right library — that shapes how the work will go.
Explore
When to use. Read-only exploration of a codebase or a document tree. “Where is X defined?” “Which files reference Y?” “Find every place we call the deprecated API.” Explore has Read, Grep, Glob, and a few adjacent tools, but no Edit or Write. It cannot change your files.
How to prompt it. Tell it the question, name the search breadth (quick, medium, very thorough), and describe the expected shape of the answer.
Plan
When to use. Software-architect work. “Sketch a plan for adding OAuth to this app.” “Given this bug report and the surrounding code, propose a fix strategy.” Plan can Read but cannot Edit or Write; its job is to think, not to code.
How to prompt it. Describe the goal, the constraints, and the shape of the plan you want (step-by-step, options-and-tradeoffs, whatever fits).
general-purpose
When to use. Multi-step tasks where you genuinely want the agent to think, act, and take you home. It has the full tool set — reading, editing, running shell commands. This is the right choice when the work is small enough to fit in one delegation but too complex for one prompt in the main session.
How to prompt it. Same discipline as any subagent — enough context to work alone, an explicit definition of “done”, and a target for the summary.
claude
When to use. The catch-all. It has every tool. When none of the specialists fit and you just want another instance of the main assistant to work in isolation, this is the one.
Custom / project-specific agents
Beyond the built-ins, teams often define their own — a db-migration-reviewer agent, a frontend-a11y-audit agent, a release-notes-drafter agent. Each one is a system prompt plus an allowlist of tools plus a name. Same shape as the built-ins; scoped to your team’s work.
Picking one, in practice
The mental algorithm:
- Read-only? → Explore.
- Thinking, no editing? → Plan.
- Actually doing the work? → general-purpose, or a custom agent designed for it.
- None of the above? → claude.
Overshoot on tools and you invite mistakes. Undershoot and the agent will report back “I would need to Edit a file to answer this.”
Check your understanding
Quick check
—