Free resources · For businesses

Does telling Claude the reason for a rule make it follow the rule better?

Adding the reason to a rule cut speech-unfriendly writing, against the rule alone, by 28 percent on one model and 69 percent on another. An irrelevant reason did nothing.

Does giving Claude the reason change what it writes?

You give Claude a rule and it follows the rule badly. A reason may help. Anthropic's example gives the reason for a rule, and says "Claude can generalize from the explanation". The same guide says to think of Claude as "a brilliant but new employee who lacks context on your norms and workflows".

What was tested?

Anthropic's own example, lightly reworded. The rule is "never use ellipses". The reason is that a text-to-speech engine will read the reply aloud. We wrote ten short tasks for made-up small businesses, such as a radio spot for a bakery. Each had prices and days, and many had percentages, that tempt symbols and abbreviations. We compared four prompts:

the task alone;
the task plus the rule;
the rule plus an irrelevant reason of similar length;
the rule plus the speech reason.

We used two models, Claude Haiku 5.5 and Claude Haiku 4.5. A script counted items a speech engine cannot say well: symbols such as $ and %, and abbreviations such as St.

What happened?

Items per reply, averaged:

Haiku 5.5: 2.75 with the task alone, 2.58 with the rule, 2.65 with the irrelevant reason, and 1.85 with the speech reason.
Haiku 4.5: 3.35 with the task alone, 2.55 with the rule, 2.30 with the irrelevant reason, and 0.80 with the speech reason.

With the speech reason, 5 of 40 replies on Haiku 5.5 and 10 of 20 on Haiku 4.5 had none at all.

What does it mean?

The irrelevant reason of similar length did not match the gain. On Haiku 5.5 it was close to the rule alone, and on Haiku 4.5 its small change was not significant. Compared with the rule alone, the relevant reason cut the leftovers by 28 percent on one model and 69 percent on the other. On Haiku 5.5 the average was still 1.85 items per reply.

How far can this be trusted?

Two small models, 10 tasks and one kind of reason. We chose the tasks to tempt symbols, and we followed Anthropic's own example.
The count is a list of characters and abbreviations. It does not judge how a reply sounds when read aloud.
We tried only one irrelevant reason. Dollar signs stayed the most common leftover.

What we would do

Write each rule with its reason: who will use the result.
Give the reason in one plain sentence about the real use.

Check

Pick one rule you give Claude today. Add its reason, and compare ten replies with and without it.
This result covers two small models and one kind of reason.