Where the rules go: a field guide to instruction placement

Models don’t read prompts with uniform attention. Beginnings anchor, endings stay fresh, named structure concentrates focus, and the middle of long prose is where instructions go to be technically present. None of this is mystical - it falls out of how attention behaves over long contexts - but it means where you say something is part of what you said.

§1The attention shoreline

Compliance across a long prompt looks like a beach profile: high at the start, high at the end, sagging through the middle. The sag deepens as the prompt grows - in a 300-word prompt, position barely matters; past a thousand words, a mid-prompt rule can lose double-digit compliance versus the same rule at either edge. If a rule must survive anywhere, it goes near an edge or inside named structure. The middle is for scenery: context, background, description.

§2What goes where

• Top - identity and hard rules. Who the agent is, what it must never do. The opening frames everything after it; safety rules here color the interpretation of the whole prompt.
• Named sections - everything with authority. A rule inside # Escalation beats the same rule loose in prose. Headers act like scoping: they tell the model what a block of rules is for, so retrieval at answer time is structural, not incidental.
• Bottom - output shape and process. Format contracts, step ordering, “before you answer, check…” live best near the end, closest to the point of generation.
• The user turn - per-request instructions. Anything that changes per request (today’s date, the active document, this ticket’s metadata) belongs in the turn, not the system prompt. System prompts are for invariants.

§3Repetition is a tax, not a fix

The common workaround for the mid-prompt sag is repeating important rules at the end. It works, and it costs you: two copies drift apart under editing, and the second copy trains reviewers to skim. Repetition treats a structure problem with duplication. Restructure first - move the rule to an edge or a named section - and repeat only when a measured compliance gap forces it, with a comment marking the duplicate as a deliberate echo.

# Escalation   // compiles to a named <Escalation> section
IF message mentions legal_action THEN Agent MUST hand off TO human_agent.

# Background
@@ scenery: safe to sit mid-prompt
Our returns window IS 30 days FROM delivery.

§4A placement audit you can run today

1. List your five most safety- or money-critical rules.
2. Find each one’s position: word count from the top, and whether it lives under a named header.
3. Any critical rule past the 60% mark and outside a named section gets moved this week.
4. Re-run your sad-path tests. Placement moves are free wording-wise - if a test flips, it was position all along.