MCP will not produce the same result on every site, or from every prompt.

That does not mean the agent is unreliable. It means the agent is working from what it can read, what you asked it to do, and how you respond after the first result.

When something does not land the way you expected, start there.

Start with site structure

Before an agent changes a site, it has to understand what is already there. A structured site gives the agent clearer signals. An inconsistent site gives it more room to guess.

For example, imagine asking the agent to add a testimonial card that matches the rest of the site. On a structured site, the agent may find an existing testimonial component, spacing variables, a naming pattern, and similar sections to reference.

On an inconsistent site, it may find three different card styles, duplicate classes, hardcoded spacing, or similar-looking components built in different ways. The request sounds simple, but the structure underneath is unclear.

Common structure issues that can affect MCP output include:

• Several versions of the same pattern, like card, card-v2, card-new, and card-final
• Color or spacing values set directly on elements instead of stored as variables
• Repeated sections copied manually instead of turned into components
• CMS collections or fields that overlap, duplicate, or use unclear naming
• Page metadata handled differently from page to page

These issues do not mean MCP cannot help. They mean the agent may need to inspect first, ask clarifying questions, or make smaller changes with more review.

Before a larger workflow, start with an inspection prompt:

Review this site's collections, fields, classes, variables, components, and repeated patterns. Tell me what structure you find before I give you a task.

Break larger workflows into stages

MCP works best when a larger workflow is broken into clear stages.

That does not mean every prompt has to be tiny. It means each step should have a clear goal, a clear stopping point, and a result you can review before moving on.

For example, instead of asking the agent to "set up a blog," break that work into a sequence:

1. Review the existing CMS structure and suggest the collection fields.
2. Create the Blog Posts collection.
3. Add or import the first set of items.
4. Add any missing fields after you review the structure.
5. Connect the collection to a page or template.
6. Check the result and summarize what still needs cleanup.

The same principle applies to canvas work. Instead of asking the agent to redesign a full page, ask it to inspect the existing structure, update one section, confirm the classes and variables it used, then move to the next section.

Breaking work into stages helps the agent complete each task with better context. It also gives you a chance to review the output while the work is still easy to adjust.

Refine instead of restart

When output does not match what you intended, you usually do not need to start over. A targeted correction prompt works better than a broad redo.

Instead of:

This does not look right.

Try:

The card spacing is using hardcoded pixel values instead of the site's spacing variables. Update only the card spacing to use existing spacing variables. Do not change the layout, typography, or component structure.

Tip: If you aren't sure why the result missed the mark, you can even ask the agent to help diagnose the issue.

Try:

Based on my original prompt and the result you produced, what did you understand the task to be? What context, constraint, or instruction would have helped you get closer?

How Agent Instructions help

If you use MCP on the same site regularly, some guidance may come up again and again: naming conventions, design system rules, brand terms, component guidance, or approval habits.

That kind of guidance may belong in Agent Instructions. For one site, that means site-level instructions. If your team works across multiple sites, Shared Libraries give you a way to share the same skills and rules across a workspace.

Because instructions can be shared through a Library, they can also reference shared resources like components, variables, and assets. That gives the agent a clearer path to the design system resources it should use.

To learn more about Agent Instructions and see examples, visit the Webflow MCP developer docs.

Ready to continue?

Click Complete & continue for your final checklist, resources, and a guided activity you can run on your own site.