Use Credits Efficiently

Credits track how much work the agent does, so a few good habits go a long way. Most wasted credits go to the same place: a build that misses what you meant and has to be redone. Practicing good prompting habits keeps that from happening.

Be Specific Up Front

The surest way to spend fewer credits is to need fewer correction rounds. A vague prompt still builds something, but the agent has to guess, and you end up paying for follow-up after follow-up to add what you meant all along. A specific first prompt lands close on the first build.

Cover the purpose, the data, and the look and feel in your opening prompt:

• Specific (one focused build): "A single-window expense tracker. Each entry has a date, amount, category, and note. Show a running total at the top, and let me filter by category and month."
• Vague (several paid rounds): "Make me an expense app."

Both build something, but the second leaves the fields, the total, and the filter for you to add one paid round at a time. See Writing Effective Prompts.

Plan Before You Build

For anything substantial, use Plan mode to agree on the approach before the agent writes code. Talking through a plan costs very little, while building the wrong structure and unwinding it costs a lot, especially on multi-view apps where the foundation can use a lot of credits to change later.

Open in Plan mode with something like:

• Try: "Before writing any code, outline how you'd structure a client tracker with contacts, deals, and a dashboard, and let me confirm the approach first."

Once the plan looks right, use Build mode to action the plan. See Plan & Build Modes.

Point Instead of Describing

When something looks off, annotate the live app or attach a screenshot instead of writing a long description. A wordy explanation is easy to misread, and a misread fix is another paid round.

• Costly: "The button near the top, the blue one, no, the other one, make it smaller and move it to the right a bit."
• Cheaper: Annotate the exact button and send "Make this smaller and align it to the right."

Pointing hands the agent a precise target, so it changes the right thing the first time. See Annotate.

Keep Requests Focused

One change at a time is cheaper than a sprawling, do-everything prompt. When you bundle several unrelated changes and one goes wrong, you often re-run the whole request to untangle it. Smaller asks are easier for the agent to get right and easier for you to check.

Instead of one overloaded prompt like "Add dark mode, fix the CSV export, add a settings page, and rename the tabs," send each change on its own:

1. "Add a dark mode toggle to Settings."
2. "The CSV export is missing its header row. Include it."
3. "Rename the Home tab to Today."

You can queue these so they run back to back without waiting on each one. See Iterating.

Revert Instead of Patching

If a direction isn't working, roll back to a good version and re-prompt rather than stacking fix on top of fix. Layering corrections onto a confused build tends to compound the confusion, and every attempt costs credits. See Version History.

When you've sent three or four follow-ups and the app keeps drifting, stop and roll back to the last version that worked, then re-prompt that step more clearly:

• Try: "Starting from the working version, sort the table by newest date first, and keep the total at the bottom updated as I add rows."