Every time you tell AI “that is too generic,” “lead with the conclusion,” “challenge my assumption,” or “do not change that language,” you are teaching it something. The value is lost if you have to teach the same lesson repeatedly.

Before you start

First ask the AI what prior history, saved memory, project context, or instructions it can actually access.
If it cannot see enough history, use the manual fallback: provide selected prior work, saved instructions, or examples of corrections yourself.
Do not assume the AI can inspect your entire account history.

The exercise

Mine the lessons already there

Ask the AI to identify recurring instructions, corrections, and preferences supported by the context it can actually access.

Separate task-specific from durable

“Make this email shorter” may be one-time feedback. “Lead with the conclusion” may be a durable instruction. Keep only patterns that recur.

Group the instructions

Organize them into presentation preferences, working preferences, decision preferences, and recurring corrections.

Review the evidence

Delete anything AI inferred without support. Add exceptions where a rule is not always true.

Create the standing version

Turn the accepted patterns into a short instruction set AI can act on.

Maintain it

When you catch yourself repeating the same correction, ask whether it belongs in the standing instructions.

Copy this prompt

Paste this into your AI tool

Copy prompt

Review the prior history, saved instructions, memory, project context, or other material you can ACTUALLY access from our work together.

Identify recurring instructions, corrections, preferences, standards, and decision rules I have given you.

Separate them into:
1. Communication preferences - how I want information presented
2. Working preferences - how I want you to approach tasks and problems
3. Decision preferences - recurring criteria I use when evaluating choices
4. Repeated corrections - things I have redirected more than once

For each item, briefly explain the evidence you are relying on. Do not invent a preference because it sounds reasonable.

Then separate the findings into ONE-TIME CORRECTIONS and REUSABLE WORKING INSTRUCTIONS, and draft a concise reusable instruction set for me to review.

Review your result

Is this actually true about how you work?
Is it supported by repeated behavior or only one example?
Would following this instruction improve future work?
Is there an important exception?
Is anything too personal or sensitive to keep in standing instructions?

Operator example

Executive example

You repeatedly tell AI to give you a recommendation rather than six equal options, distinguish facts from inference, quantify economics when possible, and challenge weak assumptions. Those are not instructions for one memo. They describe how you want an AI thinking partner to work with you.

What not to do

Do not turn every preference into a permanent rule. Too many instructions can conflict, become stale, and eventually create the drift addressed in Run an AI Checkup.

Save the result

Save the approved instructions in a location you control and, where appropriate, in the standing-instruction feature of the AI tool you use.

Next resource

Portable Context Package

Make sure valuable working context is not trapped inside one account or vendor.

Go to Portable Context Package

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