Delegate a bounded review
You will define a narrow task, explicit tools and an evidence-shaped return, then verify its findings in the main session. Outputs are illustrations.
Official subagent documentation checked September 30, 2026 describes built-in Explore, Plan and general-purpose options plus custom Markdown definitions. Subagents keep task work separate; delegation does not make the returned answer correct by itself.
Decide whether a specialist is needed
A short inline review may be simpler. Delegate when searching source and draft files would clutter the main conversation, or when a reusable specialist helps. One reviewer is enough for this exercise; parallel workers and agent teams come later.
Review the supplied course-update draft against source.txt only.
Return unsupported claims, exact source passages and a corrected draft.
Do not edit, run commands, connect services, send or publish.
If either input is missing, state that limitation.
Provide actual file paths and task boundaries. A vague “make it better” invites scope drift.
Define project scope and tools
Create .claude/agents/course-copy-reviewer.md in the practice repository:
---
name: course-copy-reviewer
description: Compare course copy against the supplied source files.
tools: Read, Grep, Glob
model: inherit
---
Review only the source and draft identified by the request.
Separate supported facts, unsupported claims and missing evidence.
Return file references and exact passages for each finding.
Propose corrected text without modifying files.
Do not infer a date or completion status absent from the source.
The explicit allowlist omits Bash, PowerShell, writing and MCP. Merely denying Write and Edit while retaining a shell would not make a workflow read-only: a shell can still change files or call services.
Inspect the resolved tools in your installation. Defaults can inherit a broader pool; misspelled or unavailable tools can prevent launch. The model: inherit field follows the session model, not a free execution tier.
Invoke deliberately
Use /agents to inspect definitions and the supported invocation route in your version. Ask explicitly:
Use the course-copy-reviewer subagent on source.txt and draft.txt.
Return findings only. Keep all files unchanged.
After the review, summarize what was inspected and what remains unverified.
Illustrative return:
Finding: launch date has no source support.
Evidence: source.txt says “date not set”.
Draft claim: “launches Sunday”.
Proposed correction: “The launch date is not set.”
Coverage: two specified files; no external verification.
The subagent should not invent a finding to satisfy the format.
Verify in the main session
Read the cited passage, compare the draft and inspect any proposed change yourself. If the source has changed since the review, recheck the current file. A returned summary is useful evidence to inspect, not a replacement for it.
Missing source should produce a limitation. A second pass can test whether an accurate draft is left intact. Keep revisions separate from the review request so scope stays visible.
Add memory or isolation only with a purpose
Custom subagents can have persistent memory scopes. This exercise needs none: customer drafts and temporary launch facts should not become reusable memory by accident. Shared project memory has version-control implications.
Worktree isolation provides a separate repository copy for write tasks, but it is not a full network or external-service sandbox. A readonly reviewer does not need it here. Plugin subagents ignore some frontmatter such as hooks, mcpServers and permissionMode in the current docs; do not assume a project definition behaves identically after packaging.
Graded practice
Easy: boundary
Define two input files and the return format. Success: no editing scope.
Intermediate: tools
Inspect the allowlist. Success: no shell, writing or external-service tools.
Challenging: verification
Test wrong, correct and missing-source drafts. Success: evidence-backed findings or honest limitation.
Troubleshooting
Not discovered: check project/user location and definition format.
Cannot launch: inspect unresolved tool names before widening access.
Unexpected editing: inspect actual tools and delegation scope.
No findings: distinguish verified absence in scope from universal correctness.
Too much coordination: use one inline review until delegation saves work.