01
What Claude Code is
Claude Code is an AI coding assistant that can inspect a project, explain files, propose a plan, edit code, and run approved commands. It is useful because it can work with the project context instead of only answering a pasted snippet.
It is still an assistant. It can misunderstand requirements, change working behavior, expose secrets in logs, or report a passing command that did not test the user flow you care about. Keep a human decision at each boundary.
02
What it is good at—and what it can break
- Good at tracing a feature across files
- Good at turning a clear checklist into small code changes
- Good at explaining errors and proposing tests
- Can break routes, auth, payments, database queries, and responsive layouts
- Can make a broad refactor when you asked for a small fix
- Can infer missing requirements incorrectly
- Can say “done” when only a build command passed
03
The audit-first workflow
- State the goal and what must not change
- Ask Claude Code to inspect and report before editing
- Review the plan and unresolved questions
- Approve one small change
- Review the diff and run a focused test
- Commit a known-good checkpoint
- Repeat only after the previous step is understood
04
Safe prompt examples
Prompt 31
Audit before editing
Copy-paste prompt
Before changing any files, inspect [FEATURE]. Explain current behavior, relevant files, likely cause or gap, risks, and a small implementation plan. Do not edit yet.
Prompt 32
One change at a time
Copy-paste prompt
Implement only [CHANGE]. Keep the current design, routes, data contracts, and unrelated behavior unchanged. After editing, run the narrowest useful test and report the exact files changed.
Prompt 33
Stop and report
Copy-paste prompt
If requirements, data shape, permissions, or expected behavior are unclear, stop and ask me or write a report. Do not guess, refactor, or claim success without evidence.
05
Review what changed
- Read the diff, not only the summary
- Check for unrelated formatting or renamed files
- Look for secrets, URLs, user IDs, and changed permissions
- Test the main flow and one failure state
- Confirm database and payment behavior separately
- Ask for a list of tests run and tests not run
06
Beginner mistakes to avoid
- “Fix everything” prompts
- Letting an AI tool work without a Git checkpoint
- Pasting secrets into chat
- Accepting a new dependency without understanding it
- Treating a clean build as proof of correct behavior
- Changing database policies without tests
- Continuing after the AI says it is uncertain