Start Small

Agents do best with a narrow, well-defined task. Asking one to "improve the codebase" is like asking a new hire to redesign the company: they will change too much, too fast, with too little context.

Large, vague prompts produce large, vague results. The agent may touch dozens of files, introduce subtle regressions, and exhaust your context window before finishing. Small tasks keep diffs reviewable, costs predictable, and failures recoverable.

1. Pick one file or one function. Not one feature, one unit.
2. Define done. "Add input validation to createUser" is better than "make the user form more robust."
3. Run it once. Approve or roll back before expanding scope.
4. Stack wins. Three completed small tasks teach you more about the agent than one abandoned big one.

• The prompt is longer than a paragraph.
• The agent touches files you did not expect.
• You cannot explain what changed in one sentence.
• The task spans multiple domains (UI, API, tests, docs) at once.