Optimizing for Performance and Credit Usage

This article summarizes tactics for optimizing your work with an AI agent while keeping credit usage low.

Before you start

Familiarize yourself with terminology – Before you begin prompting, review the Glossary of Terms. This will help familiarize yourself with the terminology used and will help you understand some of the most common terminology behind an AI.

Plan your site’s structure upfront – Sketch or outline your page sections before starting the build. Decide on the number of pages, their hierarchy, and what each section should contain. A planned structure minimizes exploratory prompting and allows the agent to execute your vision efficiently. Head over to Webpage Anatomy to understand all of this better.

Write lean Project Instructions – Treat these as the agent’s long-term memory. Keep only essential brand, typography, and layout rules (under 2,000 characters) so you don’t waste tokens re-teaching context in each chat.

Pick the right AI model – Use Normal for simple tweaks, High Intelligence for larger bodies of work, such as ideating on the page styling and layout or tweaking in some complex integrations, and Max only when the problem is complex enough that it’s cheaper to solve in one pass.

Use built-in tools to optimize the spend

Use the Element Selector and Visual Editor - Use these tools for changing text, colors, padding, or margins directly in the interface. Visual Editor is available when one element is selected and does not use credits.

Use Page Prompt Builder and Annotate - Page Prompt Builder helps you select supported page sections and save them as instructions. Annotate lets you draw on the page preview and add a snapshot to the chat with ADD TO CHAT. These tools help you give precise instructions. Review the result in Preview and use Publish separately when ready.

Select multiple elements carefully - Hold Ctrl on Windows or Linux, or Command (⌘) on macOS, to select additional elements. Visual Editor appears only for one selected element. For multiple selections, use a prompt.

Reference resources instead of describing them - Use the “@” handle to link products, bundles, or surveys instead of pasting details manually. Or prompt the Agent to build out the product for you.

Upload screenshots for inspiration and clarity - When you have an example of a desired section, upload or paste a screenshot into chat. The agent can use it to understand the layout.

Prompting tactics that cut spend

Use one chat for one body of work - Start a new chat for each coherent workstream, such as a newsletter component, popup, product page, or navigation change. Start a new chat when you begin unrelated work, the current chat contains failed experiments or conflicting instructions, or the Builder recommends it because the context is getting long.

Focused chats reduce unrelated context and make it easier to find, review, restore, or continue a body of work. Project-wide rules belong in Project Instructions, so you do not need to repeat them in every chat.

Be explicit and concise - Give narrow instructions such as “Change the hero headline text only; leave spacing untouched.”

Group related edits intelligently - Combine small, related changes into one precise prompt instead of using many fragmented prompts.

Avoid conversational or emotional phrasing - Use factual, directive language so the agent can interpret your intent efficiently.

Keep prompts deterministic - Note what must not change, such as “Keep all existing IDs and class names,” to prevent changes that require correction prompts.

Versioning and recovery

Use Version History to roll back - Open Version History, select the version you want, and choose Restore to load it into the current chat. Restore does not publish the site. Use it instead of spending credits on prompts that try to undo changes.

Use New Chat when exploring alternatives - Open Version History, select a known-good version, open its three-dot menu, and choose New Chat. This creates a separate chat from that version while preserving the original chat and its history. Continue prompting in the new chat. There is no separate Fork button.

Only the most recent chat accepts new prompts. Earlier chats are marked Read-Only, but they can still be used as the source for a new branch through Version History and New Chat.

Restore and New Chat change Builder project or chat state only. When you are ready to make the chosen build live, review the preview and use the current Publish action separately.

Execution hygiene and usage details

Prefer small, testable changes - Ask for one section or component at a time, review it, and continue if it looks right. This minimizes token-heavy retries.

Check your usage - Open Your usage to go to My Plans, where usage and your remaining balance update in real time. To inspect one prompt's cost, hover over that prompt in the chat. The details show execution time, reasoning-step count, and individual dollar cost.

Use the daily free actions - With the Normal model, users receive five free publishes per day and five free small changes per day. A small change is a prompt that uses up to five reasoning steps. Dedicated Free Bug Fixes are also free. These allowances are separate from monthly plan usage and apply to the Normal model, not High Intelligence usage.

Reserve High Intelligence and Max for complex tasks - Major layout overhauls, complex animations, or multi-file restructures may need a more capable model.

Do pure styling touch-ups in the Visual Editor first - Make credit-free styling changes there before prompting for structural changes.