How to make every AI helper follow your house rules

We ruled out the long dash in our writing, told our main AI session, and still found long dashes in the work our helpers sent back. If your rules keep vanishing in a second tool's output, that tool is probably not being careless.

Our rule: a helper follows only the rules written in its own instructions. It does not inherit what you told the main session, so each rule you care about must be written inside each helper's own definition, as a direct order.

The same task, seen from two places

• Main session

  Holds your house rules, including the forbidden long dash

• Helper it starts

  Sees only its own instructions

  The rule stops here

What we assumed, and what we found

We assumed the writing rules we gave our main AI session reached every helper it started. They did not. Each helper begins with only its own instructions, and that is by design, so what you tell the main session stays in the main session.

The habit was a long dash, the punctuation mark we had ruled out. We had already written about editing machine habits out of writing. We just kept finding it in helper output.

It is an easy mistake because the main session feels like the whole system. You talk to it, it follows your rules, and it hands work to helpers you never see. From where you sit, the rules look universal. From where the helper sits, they do not exist.

The helper had never heard the rule at all, so there was nothing to ignore.

We run our AI work on Claude Code, built by Anthropic. That is the setup we are describing, and we are not claiming how any other product behaves. Check your own tools before you trust ours.

Notice that only the last row changed what came back.

Where the rule lived	What came back
Main session only	The forbidden long dash, in helper output
Mentioned once, mid-instructions	The forbidden long dash, still
A direct order in the helper's own definition	The habit stopped

Why a passing mention does nothing

In our case a rule mentioned once, in the middle of a helper's instructions, changed nothing. The forbidden habit kept coming back. Our opinion: a rule has to be written as its own plain statement.

Our advice for writing one: say what to do, in the imperative. Give one rule per sentence. Name the replacement, so the helper has somewhere to go: use a period or a comma instead. Skip the polite hedge, because "try to avoid" reads like a suggestion.

The rule was technically there, and it still did nothing. Presence is not the test. The test is whether the output changes.

Two cautions. First, do not paste your whole rulebook into every helper. Pick the rules that touch the work that helper does, and state them cleanly. Second, resist the urge to explain the history. A helper needs the order. Leave out the story of how you came to need it. Your tools may handle more context, so test before you trim.

Tap a card

• Rule told to the main session only What goes wrong The helper never sees it How to check Open the helper's own instructions and search for the rule's words

  Rule told to the main session only How to check Open the helper's own instructions and search for the rule's words
• Rule buried as an aside What goes wrong It reads as background How to check Read the instructions once; if you have to hunt for the rule, move it up

  Rule buried as an aside How to check Read the instructions once; if you have to hunt for the rule, move it up
• Rule worded as a hope What goes wrong It reads as a suggestion How to check Rewrite it as a direct order with a named replacement

  Rule worded as a hope How to check Rewrite it as a direct order with a named replacement
• Habit comes back anyway What goes wrong The output still shows it How to check Search the output for the mark you ruled out before you trust the work

  Habit comes back anyway How to check Search the output for the mark you ruled out before you trust the work

Tap each card for the check you can run yourself.

How to put the rule inside every helper

Open each helper's own definition and write the rule inside it, in plain words, as a direct order. Do this for every helper, including the quiet ones, because a helper you skip is a gap. This was the only fix that worked for us: the rule went into the helper's own definition.

Start with the rule that costs you the most when it breaks. For us that was a punctuation mark. Yours might be a forbidden claim or a required disclaimer. One rule done properly teaches you more than several done loosely.

The work is dull, and that is good. List the helpers. Add the same rule to each. Then read what each one sends back.

The routine, in order

1. List the helpers
2. Write the rule in each
3. Read the output New
4. Rule holds

That routine is how we Optymize an AI workflow. The last gate is the easy one to skip. The only proof that the rule landed is a look at the output, because a rule that exists on the page proves nothing about the work.

Before you call it done, run a pre-mortem. Ask how the rule could still fail, then check that exact spot. In our view the likeliest answer is a helper you forgot. If you would rather have a hand, see our AI automation service.

Example check of three helpers

Helper A

• Rule in its own definition Pass
• No forbidden mark in output Pass

All passed

Helper B

• Rule in its own definition Fail
• No forbidden mark in output Fail

2 failed

Helper C

• Rule in its own definition Pass
• No forbidden mark in output Pass

All passed

Keep the rule in step as you add helpers

Keep the rule alive by treating it as a copy you maintain. Give every new helper the rules on its first day. When you change a rule, change it in every helper's definition. Our view is that the copy you forget is the one that breaks.

This is the dull half of the job. The first fix is a decision. The upkeep is a habit. Add a check to the day you edit the main rule: open each helper and confirm the new wording is there. It is a small chore that saves you the slow hunt for where a habit came back from.

So here is the rule we took from all this: if a rule matters, write it in the instructions of whoever does the work. Telling the manager is not telling the team. Run the checklist below on your own helpers.

Do this on your own helpers

• List every AI helper your main session starts.
• Write your top rule inside each helper's own instructions.
• Phrase it as a direct order with a named replacement.
• Read each helper's output and search it for the forbidden habit.
• Update every helper whenever you change the rule.

Ready to ship

This post was drafted with AI tools from our own project records, anonymized, and checked by us before it published. Read our editorial standards.