Let Claude Code figure out the steps

Most people tell Claude Code exactly what to do, step by step, and sit there supervising every move.

That is the slow way to use it.

Give Claude Code a finish line instead of a path. Something measurable it can check on its own. Every test passing. All forty rows in the sheet. The file saved to the right folder.

Once you set the finish line, Claude Code works turn after turn on its own. After each turn, a separate model checks the output to confirm the job is done. Claude is not marking its own homework. If the work is not finished, it keeps going. When the finish line is hit, it stops and surfaces the result.

The skill that makes this work is writing a clear target. Vague gets vague back. "Make it good" is not a finish line. "Every test is passing and the CSV has forty rows" is.

Be honest about the trade-off: it burns through tokens. Claude is doing the work you used to sit and supervise. For a small task you could verify in ten minutes, that is probably not worth it. For a real build, a scraper, a report generator, an internal dashboard, it is.

Set the finish line, step away, and come back to a finished result.