How we restore direction
when AI drifts, guesses, or
wanders outside the boundaries

_{—Part of CSS Primer for the AI Era — Working with Existing Code}

Even with clear guidance, AI sometimes gets lost.
It may drift away from the task, invent patterns,
misunderstand the structure,
or start refactoring parts of the system we never asked it to touch.

This is normal.
AI does not have a map —
it only has the text we provide.

A simple mental model:

AI loses the thread.
We bring it back.
Collaboration is the act of
re‑anchoring.

This page teaches us how to recognize when AI is drifting,
how to reset its context,
and how to regain control of the collaboration.

AI is drifting when it:

• proposes changes outside the scope
• rewrites parts of the system we did not show
• introduces new patterns
• renames classes without permission
• restructures markup unnecessarily
• misunderstands the component’s purpose
• contradicts earlier constraints
• becomes inconsistent across messages

These are signs that AI has lost the anchor.

A simple truth:

Drift is not failure.
Drift is a signal to
re‑establish context.

When AI drifts, we stop
and restate the following:

• the task
• the boundaries
• the constraints
• the desired outcome
• what must stay the same

For example:

“Let’s reset.
Our goal is to adjust spacing only.
No renaming, no restructuring, no new patterns.”

Or:

“We are focusing on this component only.
Everything else must remain unchanged.”

A simple reminder:

Repetition is not redundancy.
Repetition is how we re‑anchor AI.

We regain control by showing less:

• the smallest relevant snippet
• the exact part we want to change
• the minimal surrounding context

We remove:

• unrelated components
• legacy code
• global styles
• noise
• distractions

A simple principle:

When AI loses focus,
we shrink the window
to reclaim the shape of the task.

We remind AI what must remain untouched:

• structure
• naming
• behavior
• semantics
• accessibility
• component boundaries
• plugin/theme boundaries
• legacy constraints

For example:

“Please keep the structure and naming 
exactly as they are.
Only adjust the CSS inside this snippet.”

A simple truth:

AI respects boundaries only when
we restate them.

AI often gets lost because
it cannot infer relationships:

• between components
• between theme and plugin
• between markup and CSS
• between legacy and modern patterns

We re‑anchor by narrating:

“This markup comes from the plugin.
The theme styles it.
Do not change the markup.”

Or:

“This wrapper is required for layout.
Do not remove it.”

A simple principle:

AI understands relationships only when
we describe them.

Before continuing,
we ask AI to restate:

• the task
• the constraints
• the boundaries
• the desired outcome
• what must remain unchanged

For example:

“Before continuing, 
please summarize your understanding of the task.”

A simple truth:

Understanding is visible
when AI explains it back to us.

Sometimes the best reset is a clean slate.

We provide:

• a fresh snippet
• a fresh description
• a fresh set of constraints
• a fresh statement of intent

For example:

“Let’s start over with a clean prompt.
Here is the component.
Here is the goal.
Here are the constraints.”

A simple reminder:

A clean prompt is often
the fastest path back to clarity.

When AI is drifting,
we anchor it with:

• the current version
• the desired version
• the minimal change needed

For example:

“Here is the current markup.
Here is the desired structure.
Please generate the minimal CSS needed.”

This prevents AI from improvising.

A simple principle:

Anchors reduce drift.
Drift reduces safety.
We anchor to protect the architecture.

If AI continues to drift,
we decline:

“This is drifting from the task.
Please stay within the original structure.”

Or:

“This introduces changes we did not request.
Let’s return to the minimal refactor.”

Or:

“This is outside the scope.
Please focus only on the highlighted part.”

A simple truth:

Declining is part of guiding.
Guidance is part of collaboration
Collaboration is how we steward the architecture.

AI gets lost because it cannot see the system.
We re‑anchor because we can.

We restore:

• the goal
• the boundaries
• the structure
• the intent
• the relationship between parts
• the clarity of the task

A simple closing thought:

AI drifts.
We re‑anchor.
That is how the work stays grounded.

• Suggested Next Reading: → When AI Misunderstands the System (How to Reveal the Missing Context)

_{Tony de Araujo —New York | Lisbon}

• Top↑
• Back to Topic ↔
• Home←