1. Speed is its job. Judgment is yours.
Claude Code can read the project, edit files, and run commands. That speed is the reason to use it. It does not know which change your users will thank you for, or which “small cleanup” will blow up a release.
Your leverage is four verbs: scope the job, specify what finished means, review the change, and decide what gets kept. A vague sentence produces a vague edit, quickly. A sharp sentence produces a sharp edit, also quickly.
2. Finished has to be something you can check
Work backward from what a person will see. “Make the export nicer” is a mood. “The invoice page downloads a CSV with the columns Date, Vendor, and Amount, and the existing invoice test still passes” is a finish line.
If you cannot tell a stranger how you would know the work was done, Claude cannot tell either. It will guess, and the guess will look confident. One casual prompt is not a test of the tool. Try the brief on a real task, then look at the result yourself.
3. Outcome, limits, and one example
Three sentences beat a page of tone. What you want. What must not change. A real file to copy. Here is a brief you can steal. It is an example, not a Cloudadorn project:
Add a download button on the invoice page.
The file is a CSV with the columns Date, Vendor, and Amount.
Do not change the API.
Follow the button that already exists on the orders page.
Done means the current invoice test passes.
“Follow the orders page” is stronger than “make it look modern.” If the work is a bug, paste the real error on the next page’s terms. A limit you do not enforce becomes a suggestion.
4. Sometimes the point is that you learn it
If the goal is to understand a codebase, do not hand over the hard part and then skim the diff. Ask it to teach. “Walk me through how a refund is saved. Ask me a question after each step. Do not change files.” You keep the judgment you were trying to build.
You can also ask for two or three approaches and the trade-off of each, and then you pick. Escaping the blank page is useful. Letting the agent pick the design, and then shipping it because the prose was confident, is how a wrong approach lands in twenty files.
5. What to try next
Take one real task. Write the finish line in one sentence a test or a person can check. Name one file that must not change. Point at one file to copy. Read the diff before you keep it.
When the task is big enough that two designs would both look reasonable, stop and plan before you build.