Preventing AI from Modifying Navigation and Design Without Permission: Why I Introduced a No UI Regression Rule
The Navigation Had Changed Without Instruction
At one point, I asked AI to add a blog article to this site. The task involved creating the article file and setting the frontmatter. I gave no instruction about navigation.
After the work was complete, I checked the site and found that the header navigation structure had changed. While adding the blog article, AI had decided that adding a link to the blog in the navigation would improve the site and made the change on its own.
There was no malicious intent. AI makes changes it judges to be improvements. But the result was that the navigation on a live site had changed without my knowledge.
Why This Is a Problem
When AI makes independent decisions to modify UI, several specific problems follow.
Changes become difficult to track
When changes are made outside the scope of what was requested, I have to verify everything that was touched. The wider the unintended changes, the more verification work is required.
Unintended effects can occur
Navigation structure changes can affect link behavior, page routing, and style application across multiple components. A change attached to a simple article-addition task can introduce the need for broader testing.
Design decisions that do not match the project’s intent can appear
AI uses general usability as its judgment standard. That does not mean the changes align with the specific design intent or priorities of the project.
Introducing a No UI Regression Rule
Based on this experience, I added the following rule to CLAUDE.md.
## UI / Navigation Rule
Do not modify existing UI, navigation, routing, or generated homepage output.
Only make UI changes when explicitly instructed to do so.I call this the No UI Regression Rule. “Regression” in software development refers to an unintended reversion to an earlier state. Here I use the term broadly to mean preventing any unintended UI modification—whether a rollback or an addition.
What Changed After the Rule Was Added
After adding this rule to CLAUDE.md, AI began asking for confirmation before touching navigation or design. The form changed to: “It may be helpful to add a link to the navigation. Should I make that change?”
Rather than making a unintended change, AI now presents the change as a suggestion. I can decide whether to approve it. The scope of what was requested stays clear, and the verification work after a session is reduced.
The Design Principle Behind the Rule
The reason this rule works is that it establishes a distinction for AI between “what I can change” and “what requires explicit instruction.”
The rule is not a blanket prohibition. It is a boundary: change what you are instructed to change, and do not change what you have not been instructed to change. This boundary controls the scope of AI’s work.
AI tends to add what it judges as improvements. Defining the scope of that judgment through rules is what keeps project intent and design decisions in my hands.
Summary
After experiencing AI modifying navigation while working on an unrelated task, I added a rule to CLAUDE.md specifying that UI must not change unless explicitly instructed. This clarified the scope of AI’s work and reduced the verification required after each session. Designing the boundaries of AI’s judgment affects not only the quality of its work but also the cost of reviewing it.