Effective prompting for website editing

B12's AI Agent acts on what you tell it. Every message you send applies a change to your website, creates a new version with a restore point, and uses AI credits based on the request's complexity. Writing clear, well-structured prompts helps you get the result you want on the first try, so you spend less time on revisions and more time building your site.

This article covers techniques for writing effective prompts, including brainstorming with the Agent, structuring your instructions, asking the Agent for clarification, uploading reference files, and queuing multiple tasks.

Brainstorming with the AI Agent

The AI Agent defaults to taking action, but you can use it to think through ideas without making any changes. Just tell it not to touch the site.

Try:

"Don't make any changes to the website. Just suggest three different style and layout approaches I can use for my blog.

Or:

"Create a step-by-step plan for adding a member portal to the site and share it with me to review. Don't implement anything until I approve the plan."

This is useful for exploring options, comparing approaches, or generating a plan before you commit. Once you've landed on something you like, follow up with a clear instruction to apply it.

Writing effective directives

The Agent responds to instructions by making changes. Vague or open-ended prompts leave room for interpretation which can produce more than you intended.

Instead of:

Can you make this section better? 

Try:

Rewrite the About section copy to be more concise. Keep the same layout and tone.

Direct instructions give the Agent a clear outcome to work toward.

Stating what shouldn't change

If you don't define constraints, the Agent may apply changes more broadly than you intended. Specifying what's off-limits before the Agent starts gives you more predictable results.

Common things worth locking down:

• Layout and section order

• Navigation and footer

• Colors and fonts

• Copy in sections you're not editing

• Existing functionality

Try:

Change the link color across the site to blue. Do not change the navigation, footer, fonts , or any other colors.

Constraints are guardrails. The more precisely you define them, the less work you have afterward.

Using the four-part prompt structure

The most effective prompts share four elements, even when written in plain language:

• Action — What should change?

• Scope — Where should it change?

• Constraints — What must stay the same?

• Outcome — What should the result achieve?

You don't need to label them. But including all four removes guesswork and gets you closer to the right result on the first try.

Instead of:

Improve the hero section.

Try:

Rewrite the hero headline and subheadline to be more benefit-driven. Apply this only to the hero section. Keep the layout, colors, and CTA button unchanged.

The first version leaves the Agent guessing what "improve" means and what it's allowed to touch. The second answers all of that before the Agent starts.

Grouping related changes

You can include multiple changes in a single prompt, just keep them related. Grouping unrelated requests in one message increases the chance that changes conflict with each other, only part of the prompt gets applied, or results are harder to review.

Instead of:

Rewrite the hero, fix the spacing across the whole site, update the CTA button, revise the About section, and change the footer layout.

Try:

Rewrite the hero headline and subheadline to be clearer and more specific. Make the CTA button copy more action-oriented.

Think of it the way a developer works: short, verifiable steps rather than sweeping instructions. Each prompt should do one thing and do it well.

Explaining why, not just what

Giving the Agent context behind a change, not just the change itself, helps it make better decisions about how to meet the goal.

Instead of:

Shorten the About section.

Try:

Shorten the About section. The current version is too long for a homepage; visitors should get the core idea in a few seconds before they scroll to services.

The second version gives the Agent a standard to meet. This is especially useful for copy edits, tone adjustments, and layout changes that involve judgment.

Asking the Agent for clarification before it acts

For complex or high-impact changes, you can ask the Agent to confirm its plan or ask clarifying questions before making any changes. This is a good way to catch misunderstandings early.

Try:

"Before making any changes, ask me any clarifying questions you need."

Or:

"Summarize what you're going to do and wait for my confirmation before applying any changes."

Showing an example when precision matters

When you need a specific tone, format, or sentence style, an example is more reliable than describing it with adjectives. You can share descriptions, images, or links to external websites.

Instead of:

Rewrite the headline to sound more professional.

Try:

Rewrite the headline to sound more like this: "Accounting support built for small business owners, not just tax season." Keep the sentence length similar.

Use examples for tone, format, length, structure, or voice, any time "you'll know it when you see it" would otherwise apply.

Uploading files

You can upload files directly in the AI chat to give the Agent reference material or content to work from.

Supported file types: Images, PDF, TXT
Maximum file size: 75 MB

Ways to use uploaded files:

• Share your logo or brand images for the Agent to place on your site
• Provide a document, such as a resume, services list, or bio, for the Agent to reference when writing or editing copy
• Upload a PDF or text file for the Agent to read and draw from

Example:

"Here is my services brochure. Use the information in this document to rewrite the Services section of my website."

Queuing tasks

If the Agent is already working on a task, you don't need to wait before sending your next instruction. You can queue additional messages, and the Agent will work through them automatically in order.

With the queue, you can:

• Reorder, edit, or remove queued messages at any time
• Pause the queue whenever you want to review results before continuing

Learn more about managing your task queue →

Restoring and restarting

If a result isn't what you expected, the most reliable fix is to restore the previous version and rewrite the prompt, not to layer corrections on top with follow-up messages.

Instead:

1. Restore the previous version in the AI chat.

2. Identify what was unclear in the original prompt.

3. Rewrite it with a tighter scope and better constraints.

4. Try prompting again.

A clean restart with a better prompt almost always outperforms patching, especially for complex changes. Version history exists precisely for this.