What does The Placement Test: Why Dr. Jonah Tebaa Says a Prompt's Structure Needs Its Own QA Pass mean in practice?
Dr. Jonah Tebaa argues that a prompt's structure needs its own QA pass because compliance failures stem from instruction placement rather than flawed wording. Since language models weight recently processed content more heavily, placing constraints early before large reference data causes them to compete for attention and fail. Through a controlled comparison on a twelve-SKU campaign, Tebaa proved that moving a banned superlatives constraint to the prompt's final line eliminated compliance breaches entirely.
Most compliance failures in AI-drafted content get filed under the wrong cause. A banned phrase slips through, and the fix that follows is almost always about wording: sharpen the rule, add an example, escalate the warning. Dr. Jonah Tebaa's recent work argues that the wording was rarely the problem in the first place. The instruction was correct. It was simply in the wrong place.
He illustrates the point with a scenario familiar to any team running AI output at volume: a paid-social campaign, twelve product SKUs, a platform ad policy that bans unsubstantiated superlative claims like "best" or "guaranteed," and a reusable prompt template already carrying that exact rule. The rule had been in the template for months. It still failed.
A Controlled Comparison, Not a Guess
Rather than treat the failure as a one-off, Tebaa had the team run a direct comparison between two versions of the prompt. Version A stated the no-superlatives constraint early, in the instructions block that preceded the product data — the conventional structure, modeled on how a brief or style guide is normally written, rules first. Version B was identical in every respect except one: the same constraint was restated as the final line of the prompt, immediately before the instruction to generate.
The results were not subtle. Version A produced banned language in two of twelve ad blocks. Version B produced none. Nothing about the rule itself changed between the two versions — only its position in the prompt.
Tebaa's explanation is a mechanical one, not a stylistic preference. He points to how these models weight the content they've most recently processed more heavily than content earlier in a long context window. An instruction sitting several paragraphs before a block of product data has to compete with everything that comes after it for the model's effective attention by the time it starts generating. An instruction sitting as the last line before the generation call has nothing after it to compete with.
This lines up with published research on how these models actually read long inputs. A Stanford-led study of long-context language models found that performance is often highest when relevant information occurs at the beginning or end of the input context, and degrades for information placed in the middle — the same positional effect Tebaa is exploiting by moving the compliance rule to the last line of the prompt.
Where Constraints Go to Get Ignored
The practical value of Tebaa's framing is that it gives teams a specific set of places to check, rather than a vague instruction to "write clearer prompts." In his review of reusable templates, the pattern repeats across a small number of predictable locations: rules parked in a system-prompt preamble and never restated near the actual task; constraints written before a long block of reference data that pushes them further from the point of generation as the data grows; behavior modeled through few-shot examples without the underlying rule ever being stated in words; and constraints set once at the start of a multi-turn conversation, then left to compete silently against everything said in the turns since.
In each case, the fix he recommends is the same: move the constraint closer to the moment the model actually generates, and for anything carrying real compliance or safety weight, put it last.
The Decision Most Teams Skip
What makes the example worth studying, in Tebaa's telling, is not the fix itself but the decision the team faced four hours before their campaign was due to launch. They had already drafted twelve ad blocks under the flawed structure and could have shipped that output after a manual proofread — a roughly twenty-minute task, and one that leaves the compliance check resting on a tired human reader catching what the model had already missed once. Or they could discard the batch, regenerate with the corrected structure, and spend roughly fifteen minutes doing it. They chose to regenerate, and then went further: they rewrote the template's default structure so every future SKU batch would carry constraints in the position most likely to hold.
Their proofreading step changed after that, from a full read-through of every block to a spot-check of a handful — a meaningfully lower-risk position at the same production volume.
The broader argument Tebaa is making to the executives who read his work is aimed less at prompt-writers than at whoever signs off on AI templates for scaled use. A template that reads correctly is not the same as a template that performs correctly, and the gap between the two shows up specifically at the moment volume goes up and nobody is reading every output anymore. Treating a prompt's structure as a one-time design decision, fixed once and reused indefinitely, is what lets that gap persist. Testing structure the way a team tests subject lines or landing pages — deliberately, and again whenever the stakes or the length of the prompt change — is what closes it.
Tebaa's wider body of work on applied AI implementation, including his framework for classifying tasks before writing a single prompt, is available on his blog at jonahtebaa.com/blog.
