Tell It What to Leave Alone

---

Intro

In B-08 you learned what to do when the AI goes wrong: start over, rewrite from scratch. Here is how to stop it going wrong in the first place. Prevention costs one line. Correction costs a whole extra conversation.

---

The concept

The AI will do what you asked — and everything it thinks follows naturally from that. If you ask it to rewrite an email, it may also reorganise your paragraphs, adjust your greeting, add a summary sentence at the end. None of that was in the brief, but all of it seemed like part of the job.

A constraint tells it where to stop.

Without constraint:

Rewrite this project update to be clearer.

The AI rewrites. It also shortens two sections you needed in full, changes your sign-off, and adds a bullet list you didn't ask for. Technically, it did what you asked. You are now fixing things that were already fine.

With constraint:

Rewrite this project update to be clearer. Only change the wording — don't restructure the sections, don't add anything I haven't mentioned, and keep the sign-off exactly as written.

Same verb. Same task. The second version tells it where its authority ends.

---

The exercise

Part 1: Constraint surgery

Take either of these from earlier in the course:

• The opening seed you wrote in B-10
• Your persistent instructions from B-09

Or use any message you have sent recently that produced something you had to clean up.

Send it twice: once without a constraint, once with one.

Without constraint:

[Verb] [the specific thing]. [Stopping condition].

With constraint:

[Verb] [the specific thing]. [Stopping condition]. Don't [what to leave alone].

Copy these starters. Fill in your own task. Send both. Then look at what the constraint actually changed.

You are not trying to get a better response. You are watching what the constraint prevented.

---

Part 2: The constraint slot

Every well-formed message has the same structure:

[Verb] [what you want]. [Stopping condition]. Don't [what to leave alone / where to stop / what not to change].

The constraint slot — the "Don't" — can do several different jobs:

Scope: Limits where the AI works.

Only the first paragraph — leave the rest as-is.

Style: Protects how something sounds.

Don't change the tone — this is intentionally informal.

Additions: Stops the AI going beyond the brief.

Don't suggest anything I didn't ask for.

Format: Keeps the shape you already have.

Don't add headers or bullet points.

Voice: Keeps the writing yours.

Keep my sentence structure — only change the words I've marked.

You do not need all of these at once. Pick the one that protects what matters for this particular task.

---

Common constraint patterns

Keep these. Use them as a starting point when you are not sure how to word a constraint.

• "Don't change anything I haven't mentioned."
• "Only [specific section] — leave the rest as-is."
• "Don't add [headers / bullet points / a summary / recommendations]."
• "Stop after [condition: one pass / three suggestions / the first section]."
• "Keep my [voice / structure / examples] — only change [the specific thing you named]."

---

What good looks like

The response addresses exactly what you named and nothing else. The parts you did not mention are unchanged. You do not need to undo anything.

---

If the AI ignored your constraint

It does happen. When it does, the response itself usually tells you why — the AI may have interpreted your task as requiring the thing you said not to do, or the constraint may have been too vague for it to apply.

Do not patch it. Start a new message. Make the constraint more specific: instead of "don't change the structure," try "keep the same three paragraphs in the same order." Name what you want preserved as precisely as you named what you want changed.

---

Next

The verb sets what the AI does. The constraint sets what it leaves alone. E-03 covers the third piece: the context that makes both of those work — the information you give the AI before it starts.