Does it matter where you put the rule: start or end of a long prompt?

The question

We gave a passage and a strict format rule (exactly 3 bullets, each 12 words or fewer). Does the rule work better before the passage or after it?

What we found

• Haiku: 83% to 70% (no clear difference)
• Sonnet: 57% to 57% (no clear difference)

"Clear" means the two 95% ranges below do not overlap. When they overlap we say there is no clear difference, even if the percentages differ. With about 30 answers per cell, small gaps are noise.

Model	How we asked	Passed	Rate	95% range
Haiku	Rule first, then the passage	25 of 30	83%	66% to 93%
Haiku	Passage first, rule last	21 of 30	70%	52% to 83%
Sonnet	Rule first, then the passage	17 of 30	57%	39% to 73%
Sonnet	Passage first, rule last	17 of 30	57%	39% to 73%

What counts as a pass: exactly 3 bullet points and every bullet is 12 words or fewer.

What we saw reading the answers

• Almost every answer had exactly three bullets (119 of 120). The failures were bullets of 13 to 15 words instead of 12 or fewer, for example "12 / 12 / 15 words".
• Putting the rule first or last made no clear difference for either model.
• Haiku kept to the 12-word limit more often than Sonnet (46 of 60 against 34 of 60). The ranges only just separate, so treat that as a hint, not a finding.
• What to do: if a word limit matters, state it plainly and check it with code. Models count words loosely.

The exact wording

Shown with the first question as the example (the long passage is cut short here).

Rule first, then the passage

System prompt: You are a helpful assistant.

Summarise the passage below in exactly 3 bullet points. Each bullet must be 12 words or fewer. Output only the bullets.

Passage:
[passage of 176 words]

Passage first, rule last

System prompt: You are a helpful assistant.

Passage:
[passage of 176 words]

Summarise the passage above in exactly 3 bullet points. Each bullet must be 12 words or fewer. Output only the bullets.

How we tested

• Models: Haiku (claude-haiku-5-5); Sonnet (claude-sonnet-5-5). Claude models only. We have not tested ChatGPT, Gemini, Grok or others, and results can differ.
• 10 different questions, each asked 3 times per way of asking, so the answer's randomness is included.
• Every answer was scored by a short program, not by a person's opinion. We then read a sample by hand.
• Each question was sent on its own through the Claude command-line tool, with tools, memory, settings and conversation history switched off, and no web access or examples. That tool still adds a short fixed note about the machine (operating system, working folder, the account's email domain), about 900 tokens in all. It is identical for both ways of asking, so comparisons stay fair, but absolute numbers may differ a little from a plain API call. We redacted the owner's company name from one published answer where a model repeated it.
• Run on 2026-10-26. Model versions change; treat results as a snapshot.

Limits

Ten passages of about 200 words from our own guides. Longer documents may show a bigger gap.

Raw data

Every question, answer and score is in the raw data file (filter by "exp": "constraint-placement"). Re-run it yourself with the harness in the project's testlab folder, or paste the prompts above into your own account.