Quick answer: Start Claude Code work with a clean goal, a bounded repository context, a plan, small changes, and tests. Never grant broad permissions without understanding the environment and the consequences.
Before you start
Read the project’s contribution notes, check the working tree, identify the test command, and avoid exposing secrets. Create a branch or otherwise preserve a recoverable state.
The inspect-plan-edit-test loop
- Ask the tool to inspect relevant files.
- Request a short implementation plan.
- Make one narrow change.
- Run tests and inspect the diff.
- Repeat only after the change is understood.
Review generated code
Check input validation, authorization, error handling, logging, dependency changes, and test coverage. Generated code can be plausible while still being insecure or incompatible with the project.
Practise on a small project
Use a small, non-sensitive repository and document what the agent changed. Pair this tutorial with the AI agents explainer and automation project guide.
Explore your next step
Learn agentic building with guardrails
Review the ISS AI & Agentic Systems programme.
View AI curriculum →Editorial scope
Check the current official product documentation before using a CLI, model, or permission setting.