Avatar

Say why, and stop shouting

TL;DR

  • Capitals and "IMPORTANT" were a fix for older models. The newer ones over-apply them.
  • "Never" is read as having no exceptions. If there is one, write it.
  • Give each rule its reason. The AI then knows where the rule stops.

For a while, the way to make an AI follow a rule was to shout it. Capitals. "IMPORTANT." "You MUST." "NEVER do this." People still copy those prompts around.

It made sense once. The older models drifted, and a loud rule held them in place.

The newer ones do not drift. They listen. So the shouting now does harm.

What the shouting does now

Anthropic's best practices page for its coding tool says it directly: "Older models sometimes needed the emphasis. Current ones follow a plainly worded line, and piling on emphasis mostly makes Claude over-apply the rule" (Best practices for Claude Code, as of November 2026).

Over-apply means the rule spreads into places you never meant. Say you wrote "NEVER change the price on the page" because a draft once changed it by mistake. A model that takes this as absolute may refuse to fix a price you asked it to fix. It may refuse to add a new price for a new service. The rule was about one mistake, and now it governs everything near it.

There is a second cost. When every rule is loud, none of them stands out. The capitals stop carrying any meaning.

"Never" means never

The same habit shows up in quieter words. "Never", "always", "under no circumstances".

Anthropic's guide to its newest model says: "The model treats an instruction that contains 'never' as absolute. For example, write 'ask for approval before deleting files' instead of 'never delete files' if you want the model to keep responsibility over an action" (Prompting Claude Fable 5.2, as of November 2026).

So before you write "never", ask yourself one question. Is there truly no case where this is fine?

If there is a case, write it down. "Do not change a price unless I ask you to in this message." That one sentence does what the shout was trying to do. It also leaves the door open where you want it open.

Give the reason

A rule with its reason works better than a rule alone. Anthropic's prompting page makes the same point: say why, and say what to do rather than only what not to do (Prompting best practices, as of November 2026).

The reason does real work. It tells the AI where the rule stops.

Here is one of my own rules, as I give it to an AI that writes for my clients' sites:

Do not promise a reply time. One busy person answers the enquiries, and a promise they cannot keep loses the customer it was made to win.

Without the reason, the AI might also strip "we open at 8" from the page. That is a time, after all. With the reason, it knows the rule is about promises to a waiting customer. Opening hours are fine.

The reason also helps you. Six months from now you will read the rule again. If the reason no longer holds, you can delete the rule with confidence.

Old warnings pile up

If you keep a standing set of instructions for an AI, warnings collect in it. Each one was added after a bad day. Many of them outlive the problem.

Anthropic's guide warns about this too: "Memory files, notes, and instruction files collect cautions over time, and the model might have written some of them itself. Before you migrate, delete cautions whose reason no longer holds, and rewrite absolute rules with their scope and reason" (Prompting Claude Fable 5.2, as of November 2026).

I found this in my own setup. My main instruction file had grown to 598 lines. Twenty two of them began "Do NOT". Each had a reason, but the reasons were buried in history. The file is now 116 lines, and none of them shout. Prune your instructions tells that story in full.

When words are not enough

Some rules cannot bend at all. Do not read the file that holds the passwords. Do not send an email without me.

For those, a sentence is the wrong tool. Anthropic's best practices page says: "If you keep adding emphasis to a CLAUDE.md line and Claude still misses it, that's a sign the line belongs in a hook" (Best practices for Claude Code, as of November 2026). A hook is a small script, set up in the tool itself, that blocks an action every time.

Most AI tools have something like it: a setting, a permission, an off switch. Put the rules that cannot bend there. Then the words in your brief can be calm, and they only have to guide.

The short version

Write each rule the way you would say it to a sharp new colleague on their first day.

Say it once, in a normal voice.

Say why.

If there is an exception, say that too.

Related reading

Contact

Tell me what you’re working on.

Say what you are trying to fix and I will give you a straight read. If it is easier to talk, pick a time that suits you.

Book a time to talk

Prefer to write?

Shei Goldberg

I read every message myself, and I will get back to you.

We will text or call you on it.