When does it pay to repeat key instructions near the bottom of the user turn instead of trusting the system prompt?

Attempt it

Two reasons. First, recency: the bottom of the prompt is the privileged recency slot, immediately adjacent to where the model starts generating; instructions there get the highest effective weight. Second, lost-in-the-middle: when the conversation is long, the system prompt is at the top, the retrieved chunks fill the middle, and the user turn is at the bottom. A single critical rule (output format, refusal policy, citation style) repeated at the bottom punches through the middle trough. The price is a small token cost and a small risk of contradicting yourself if the two copies drift, so it is reserved for one or two truly load-bearing rules, not the whole policy.

Your answer

The position of an instruction in a prompt affects how much weight the model gives it. **Recency**, the position closest to where generation starts, is the strongest slot. **Primacy**, the position at the very top, is the second strongest. The middle of a long prompt is weakest, an effect called **lost in the middle**. In a typical RAG or agent prompt, the system prompt is at the top (primacy slot), the retrieved chunks fill the middle (the weak zone), and the user turn is at the bottom (recency slot). A critical rule stated only in the system prompt has to compete with everything in between to reach the model's attention when it generates. Mirroring a load-bearing rule at the bottom of the user turn solves this. The repeated rule lands in the recency slot, immediately adjacent to the generation start. The model is essentially incapable of ignoring it because it just read it. This trick is used for one or two critical rules: an output format constraint, a refusal policy, a citation style requirement. The cost is a small number of tokens and the risk of the two copies drifting if you only update one. So reserve the repetition for truly load-bearing rules, not the whole policy. Most teams find this works best in long contexts where lost-in-the-middle is most pronounced.

Position in a prompt affects effective instruction weight. **Recency** (closest to the generation start) is the strongest slot; **primacy** (top of the prompt) is the second strongest; the middle is weakest, the lost-in-the-middle effect. Mirroring a load-bearing instruction at the bottom of the user turn is a deliberate use of the recency slot, and it pays off in two specific situations. **Long contexts.** When the system prompt is at the top, retrieved chunks fill the middle, and the user turn is at the bottom, a rule stated only in the system prompt has to compete with everything in between. In a 32,000-token prompt with the system at position 0 and the user turn at position 31,000, the model attends much more strongly to recent content. A critical rule mirrored at the bottom of the user turn punches through the lost-in-the-middle trough and lands in the strongest attention slot. **Critical-correctness rules.** Some rules cannot be allowed to fail occasionally: an output-format constraint (always return JSON), a refusal policy (never share personal data), a citation requirement (always cite source for factual claims), a structure requirement (start with a one-sentence summary). For these, the operational cost of one occasional miss is large enough to justify a small token investment in repetition. The repetition is **selective**. One or two load-bearing rules, not the whole policy. Mirroring the entire system prompt at the bottom defeats the purpose: nothing stands out and the model treats it as noise. The price of repetition is also real: a few extra tokens per call and the risk of the two copies drifting when one is updated and the other forgotten. Production stacks treat the mirrored rules as a managed list with a single source of truth. This pattern interacts well with prompt caching.

Quick reference

TL;DR

Mirror one or two critical rules at the bottom of the user turn to exploit the recency slot and punch through the lost-in-the-middle trough, while keeping the bulk of the policy in the cached system prompt.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Imagine a meeting where you spent 20 minutes laying out the agenda at the start, then talked through a long set of slides, then asked the team to act. The agenda is at the top, the slides are in the middle, and the action ask is at the bottom. By the time people are about to act, the slides are fresh in their minds and the agenda is fading. So when you really need them to remember one specific rule, you say it again right before the action. A prompt works the same way. The system prompt is the start of meeting agenda. The retrieved chunks are the slides. The user turn is the action ask. Repeating the critical rule right before the model generates is the equivalent of saying "and remember, this one thing matters most" right before people start working.

Key concepts

Context Engineering Agents

Detailed answer & concept explanation

~7 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Sign in for 5 free unlocks/day or go Pro for unlimited. See Pro pricing.

Quick scan · 2–4 min

Everything important, quickly.

Define the recency and primacy slots, plus the lost-in-the-middle trough. Explain why a long context dilutes a system prompt only rule. Identify the two situations where mirroring earns its tokens: long contexts and critical-correctness rules. Cover the discipline (selectivity, source of truth, cache awareness). Close with the production examples in 2026 stacks.

Real-world examples

Real products, models, and research that use this idea.

Anthropic's prompt engineering guide for Claude 4.7 recommends restating critical rules near the end of long prompts.
Cursor agent mode mirrors the 'only modify files in scope' rule at the bottom of every tool-call prompt.
OpenAI's RAG cookbook in 2026 shows mirroring of citation-style rules at the bottom of the user turn.
LangChain RAG chains expose a 'reminders at end' parameter that operationalizes this exact pattern.
Enterprise compliance assistants (legal, medical) commonly mirror refusal-policy rules at the bottom in addition to the system prompt.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QHow does this pattern interact with provider-side prompt caching (Anthropic, OpenAI, etc.)?

Sign in or go Pro to see the approach for this follow-up.

QWhat is the right way to choose which rule to mirror when you have many candidates?

Sign in or go Pro to see the approach for this follow-up.

QDoes this pattern still help in 2026 long-context models that claim to handle million-token inputs?

Sign in or go Pro to see the approach for this follow-up.

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Treating instruction repetition as a sign that the system prompt failed. The repetition is a deliberate use of the recency slot, not a workaround for a broken system prompt.

Sign in or go Pro to see all red flags and common mistakes. See Pro pricing.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.