Use one loop for every bounded task
Begin with repository orientation and the user-visible outcome. Ask the agent to identify the relevant seam, then implement narrowly, run proportionate checks, inspect the diff, and report evidence. This structure works better than a giant prompt containing every possible instruction. It is also what makes it practical to run two agents in parallel once the loop is reliable.
Separate discovery from risky action
Read-only inspection can be broad; mutation should be narrow. Deployment, billing, destructive data changes, and external publication deserve explicit authorization even when implementation is complete.
- State the source of truth.
- Name prohibited cleanup.
- Keep acceptance criteria observable.
- Stop when new authority is required.
Leave a resumable handoff
Record what changed, what passed, what did not run, and which decision remains. CmdBrief keeps that session attached to the product workspace so the next run starts with evidence rather than memory.
Questions
What makes a good Claude Code task?
A clear outcome, narrow scope, explicit constraints, and commands or observations that prove completion.
Still unsure? Write to [email protected] or use the contact form. Contact CmdBrief →
Primary sources
Provider terms and product behavior can change. Confirm plan, pricing, model, and permission details in the current official documentation.