AI Instruction Sets for Product Managers and Designers

Three tells. You asked for options and got rewordings of one idea. You approved something in under a minute and could not say what made it good. You keep finding the assumption you missed two weeks later, as rework. If none of the three land, you do not need this yet.

A versioned library gets closer to standing context as it gets richer, but it still depends on retrieval. You have to remember to fetch it, and the day you are moving fastest is the day you will not. The difference is whether your standard applies by default or by discipline.

Yes. That essay argues the file you build to work with AI says more about how you work than any résumé, and that the trace it leaves on your work is what someone will eventually read. This is where you write the file, against your own specs and rationales. The essay is the argument. The afternoon is the practice.

It should encode team and product standards, not personal taste. Writing them down surfaces exactly where your stakeholders disagree, which is worth having on paper regardless. The set becomes a forcing function for alignment rather than a preferences file.

Yes, if you encode style. No, if you encode standards and constraints, which narrow the space without dictating the answer. Telling the model what you always reject is a very different instruction from telling it how to write.

That depends entirely on your rejection rate. Accept everything it proposes and they are the model's. The cutting is the authorship, which is why we do the cutting out loud in the room. One caution: feed it work you actually wrote, or it extracts its own defaults and hands them back as yours.

When your standards are stale. When it entrenches a blind spot you did not know you had. When the task genuinely needs an outside view. When the set gets so long the important constraints get diluted. We cover all four, because a method with no failure conditions is not a method.

Probably not yet. We start from the assumption that you generate real work with AI weekly and have hit the plateau. It is also not for AI engineers after evaluation frameworks, or anyone wanting an autonomous agent. The method assumes a person stays in the loop by design.

No code, no terminal, no API keys. Bring any major AI tool you already use. We build in one during the session and the method transfers to whichever you work in.

The method does not require it. You can build your set against sanitized or public artifacts and apply it to sensitive work afterward. Bring whatever your policy allows, and if you are unsure, bring the sanitized version.

Roughly fifteen minutes every few weeks, and the last block covers the loop that keeps it that cheap. The set that never gets updated goes stale, which is the most common way this fails.

Cheerfully, yes. That is the point. The Double Diamond has been the best available method for structuring judgment for decades, and almost nobody is applying it to AI work. The extra step is installing the first diamond so it runs by default.

A working instruction set running against your real work, a completed gate map, a before-and-after on the artifact you brought, the five-part starter kit, and the session recording.

Yes, and most do. Maven has a reimbursement request page and a team discount of 20% or more. Ask me for an invoice if procurement needs one.

You can access recordings whenever you need to.