A 40-page RFP review ignores the output template stated at the top — how do you fix it?

## Reading the symptom correctly 'The model ignores the template' can mean several different things, and the fixes diverge. Establish which one you have before touching the prompt. - **Position failure.** The instruction is unambiguous and the model follows it on a short input, but drifts on the long one. Compliance degrades as the document grows. - **Ambiguity failure.** The template is under-specified, and the model is filling gaps plausibly. Compliance is bad even on a two-page input. - **Conflict failure.** The document itself contains a competing format — the RFP shows its own response layout — and the model follows the nearer, more concrete example instead of your instruction. Only the first is an ordering problem. The cheap discriminator is to run the same instruction against a truncated document: if compliance is fine at two pages and collapses at forty, position is implicated. If it fails at both, reordering will not save you. ## Lever one: reorder Move the output contract so it is the last thing before generation. This costs nothing in tokens and is usually the largest single improvement, because the recency slot is the most influential position in the prompt. The cost is not zero in other senses. The prompt's shape changes, which matters if anything downstream depends on a stable prompt structure, and it can make the prompt harder for a human to read, since the reader now meets forty pages before learning what they are for. It also removes the framing benefit: a model that reads the document already knowing it must extract nine specific sections reads it differently from one that discovers the task afterwards. ## Lever two: restate Keep a compact framing instruction at the top — 'you are reviewing an RFP; you will produce a compliance response using the client template' — and put the **full, literal output contract** at the bottom. This is the instruction sandwich. It buys both the framing benefit and the recency benefit. The cost is a duplicated block on every call. For a template of a few hundred tokens against a 40-page document of perhaps 25,000 tokens, that is roughly one percent overhead — trivially worth it if it converts a 60 percent compliance rate into a 95 percent one. It stops being trivial when the instruction is itself enormous, or when the restatement is repeated many times through the body, which is a common overcorrection: interleaving the instruction every few pages fragments the document, multiplies cost, and rarely beats one good restatement at the end. One detail matters: restate the contract *verbatim*, not paraphrased. Two slightly different versions of the same rule inside one window is a contradiction the model has to resolve, and it may resolve it by splitting the difference. ## Lever three: shrink what sits between The underlying problem is the mass of low-influence material separating the instruction from the output. Reducing it helps regardless of ordering — selecting the sections of the RFP that actually bear on the response, or processing the document in sections and assembling the results, both shrink the distance the instruction has to survive. This is more work than reordering and changes the system's shape, so it is the right move when the position sensitivity is severe rather than marginal. ## Making the constraint checkable A formatting contract that a program can verify is worth far more than one only a human can judge. If the template maps to a schema, ask for structured output and validate it; a violation then becomes a caught error and a retry rather than a silent defect discovered by a client. Position bias does not disappear, but its blast radius shrinks from 'wrong document shipped' to 'one retry'. A related trick is asking the model to emit the contract's key constraints before the body of the answer. That forces the constraint into the model's own recent context, where it exerts maximum influence over what follows. It costs a few output tokens and is worth trying when reordering alone is not enough. ## Proving the fix Build a fixed set of representative documents — twenty to fifty is normally enough — and score template compliance programmatically: required sections present, order correct, headings matching, no extra prose. Compare the original ordering against each candidate fix on the same set at the same temperature. Report the compliance rate, not an anecdote. Placement effects tend to be large enough that a set of this size resolves them clearly, and having the number lets you decide whether the token cost of restating is justified rather than arguing about it. ## What not to do Do not respond to a position failure by escalating the language — ALL CAPS, 'IMPORTANT', threats of consequences. It occasionally helps a little, it is unmeasurable, and it leaves the actual structural problem in place. Do not scatter the instruction throughout the document. And do not conclude the model is incapable of following the template before you have tried it on a short input.