Use approved issue context and delegated review without external side effects.
Coding & development · Documented · Documentation checked 2026-10-04
Claude Code in terminals and supported development surfaces; Claude chat is not a local shell.
Authentication, installed version, surface, subscription or API billing, and organization policy matter.
An illustrative recipe. This website does not run the automation.
The goal and the right kind of system
Use approved issue context and delegated review without external side effects.
An assistant can review a known diff; a subagent is useful only for a bounded independent inspection. Hooks should run a known approved check, not publish changes.
Inputs and connected services
Inputs: sample diff, approved issue identity, trusted read-only tool, and review criteria.
Services: Claude Code plus approved MCP / plugin tool; optional reviewed hook or subagent.
The steps
An illustrative sequence to adapt and test before using real data or enabling ongoing work.
- Inspect the extension and its external scope.
- Read the intended issue and confirm identity.
- Perform a bounded read-only review or delegated analysis.
- Reproduce important findings and inspect any hook activity.
Outputs and human checkpoints
Outputs: supported findings, source issue reference, and a verification list.
Human checkpoints: approve extension scope and validate findings before any fix.
Exceptions, permissions, and stopping conditions
Permission boundary: no tracker comments, status updates, code edits, or outbound communication.
Stop on untrusted instructions, missing source identity, shared-resource conflicts, or unexpected hook side effects.
Adaptable prompt
Read [approved issue] and review this diff without changes. Give reproducible findings and mark hypotheses. No external writes, comments, or status changes. Report tool identity and any hook activity.
Related reading
MCP, plugins, and tool trust
Add narrowly scoped external capabilities and inspect the extension before loading it.
Subagents and hooks
Separate delegated reasoning from lifecycle-triggered automation and shared side effects.
Permission modes and stopping rules
Use enforceable tool policies rather than relying on prose instructions as access controls.
Designing an AI workflow
Make a useful task repeatable with clear inputs, steps, outputs, exceptions, and human checkpoints.
Permissions and meaningful approvals
Decide what a system may read, what it may change, and when someone needs to authorize a concrete action.
Sources and verification
Documentation checked 2026-10-04. Examples are illustrative unless described otherwise.
Features and extension relationships · Anthropic
Use plugins in Claude · Anthropic
Configure permissions · Anthropic