AI can gradually move away from the instructions, context, or boundaries you thought you had established. Old priorities remain in context, new instructions conflict with old ones, exceptions pile up, and AI may start making assumptions you never intended. That is drift. A short periodic checkup can catch it.

Before you start

Ask the AI what saved instructions, memory, project context, or other standing material it can actually inspect.
If it cannot inspect them directly, gather the instructions, context files, governance guide, voice guide, or other material you want checked.
Use a few representative tasks where you already know what good output looks like.

The exercise

Re-state the intended boundaries

Identify what AI should always do, never do, and always leave to a person.

Check instruction drift

Look for contradictory, duplicate, vague, or obsolete standing instructions.

Check context drift

Find projects, priorities, roles, facts, or assumptions that are no longer current.

Check judgment drift

Look for areas where AI is making stronger recommendations, assumptions, or decisions than the operator intended.

Test known tasks

Run a few representative tasks and compare the output with the expected standard.

Clean and reset

Remove stale information, reconcile conflicts, restate important guardrails, and save a clean version.

Schedule the next check

Review quarterly, after major role or strategy changes, or whenever you notice recurring corrections returning.

Copy this prompt

Paste this into your AI tool

Copy prompt

I want to run an AI checkup on the standing instructions and context I use with you.

First, tell me exactly which saved instructions, memory, project context, or other standing material you can ACTUALLY inspect. Do not imply access you do not have.

Then review the material available to you, and anything I provide, for:
1. outdated context
2. conflicting instructions
3. duplicated rules
4. vague instructions that could be interpreted inconsistently
5. assumptions that appear to have become embedded without explicit approval
6. areas where your recommendations or decision latitude may have expanded beyond the intended boundary
7. important guardrails that are missing or no longer prominent

Separate findings into KEEP, UPDATE, REMOVE, and CLARIFY.

Do not rewrite the final instructions until I approve the findings.

Review your result

What information is no longer current?
Which instructions conflict?
What do you keep correcting again?
Has AI begun assuming authority you intended to retain?
Are important guardrails still visible and specific?
Did a recent business or role change make older context misleading?

Operator example

SMB example

A manager originally instructed AI to prioritize rapid growth. Six months later the business is protecting cash and margin, but the old growth instruction remains in a project. AI keeps recommending aggressive spending. The answers are not “broken”; the context is stale. The checkup catches the drift.

What not to do

Do not treat AI’s self-review as sufficient verification. You decide what the instructions and boundaries should be. The AI can help identify conflicts; it should not quietly redefine the rules.

Save the result

Keep a dated clean copy of the approved instructions and context. Record major changes so future drift is easier to spot.

Next resource

Back to the toolkit

Return to the AI Operator Toolkit and choose the resource that matches the next problem you are trying to solve.

All resources

Working on this with a team?

The toolkit is built for individual use. Moving a whole team or company is a different level of work, and it is where most AI efforts stall.

See LBM AI Training

See how the full path works