Why does “act as an expert” no longer improve your output?
Because the instruction adds no information the model did not already have. A role line tells the model who to be. It does not tell the model what your reader knows, what you are optimising for, or what a wrong answer looks like. Modern models fill that gap with averages, and averages read as generic.
Role prompting, defined: opening a prompt with an identity instruction such as “you are a senior marketing strategist with 15 years of experience” in order to shift the model's tone and expertise. It is a persona instruction, not a context instruction.
The distinction matters because it predicts when the technique works. A role helps when it genuinely changes something concrete: the vocabulary used, the audience addressed, or what gets left out. A radiologist and a patient-education writer describe the same scan differently, so that role is doing work.
A role does almost nothing on classification, extraction and factual question-answering, where there is one correct output and no stylistic choice to make. Telling a model it is a world-class analyst does not change which number it pulls out of a table.
There is also a mechanical reason the old openers faded. Reasoning-capable models such as Claude's extended thinking mode and Google's Gemini reasoning tiers already run internal chain-of-thought whether you ask for it or not, so “think step by step” and “you are a brilliant reasoner” duplicate behaviour that is already switched on.
What changed in the models between 2023 and 2026?
Two things. The tricks entered the training data, and the default models got substantially better at the tasks the tricks were compensating for. A prompt technique only creates advantage while it is rare, and none of the 2023 openers are rare now.
The clearest recent marker is at the free tier. OpenAI announced on 6 August 2026 that GPT-5.6 Luna became the default model for Free and Go users with unlimited text chats, and reported that responses containing at least one factual error were 62% less common than with the previous default, GPT-5.5 Instant.
Read that as a prompting fact, not a product fact. When the baseline model is materially more accurate, the marginal value of a clever opener collapses, and the marginal value of supplying real context rises.
Anthropic now frames the discipline as context engineering: deliberately curating what the model sees on every single inference call, rather than searching for a magic sentence. That reframe is the practical shift. You stop asking “what wording unlocks a better answer” and start asking “what does this model not know that a good human collaborator would.”
One caveat worth stating plainly. Persona lines are not banned and they are not harmful. They are simply low-yield, and on long prompts they consume context window space that a concrete constraint would use better.
What are the five context blocks that actually change output?
Five blocks carry almost all the improvement in 2026: task, audience, source material, format contract, and negative constraints. A prompt containing all five outperforms a persona-led prompt on nearly every business writing and analysis task, because each block removes a specific guess the model would otherwise make.
The five blocks, and the guess each one removes
--- Task: one sentence naming the artefact and its purpose. Removes the guess about what you actually want produced.
--- Audience: who reads it, what they already know, what decision they make afterwards. Removes the guess about reading level and framing.
--- Source material: the real notes, transcript, data or draft, pasted in. Removes the guess about facts, and is the single highest-yield block.
--- Format contract: length, structure, and the exact shape of the output. Removes the guess about presentation.
--- Negative constraints: what must not appear. Removes the failure modes you already know this model has.
Negative constraints are the block most people skip and the one that most reliably fixes repeat annoyances. If a model keeps opening with “In today's fast-paced world”, one line saying that phrase is forbidden ends it permanently, where twenty rounds of “make it punchier” will not.
Order matters more than most people expect. Put the instructions before long pasted material rather than after it, so the model reads the task with the source in view rather than reconstructing intent at the end of a wall of text.
What does the replacement prompt actually look like?
It looks like a short structured brief with labelled sections rather than a paragraph of instructions. Labels, whether written as XML-style tags or plain capitalised headers, help the model separate your instructions from your pasted material, which is where most confused outputs come from.
Try this prompt. Copy it, replace the bracketed parts, and keep it as your template.
<task>
Write a [ARTEFACT, e.g. 300-word client update email] whose job is to [PURPOSE, e.g. get sign-off on the revised timeline without reopening the scope debate].
</task>
<audience>
[WHO READS IT, e.g. a non-technical marketing director who approved the original timeline]. They already know [WHAT THEY KNOW]. They do not know [WHAT THEY DO NOT KNOW]. After reading, they need to [DECISION OR ACTION].
</audience>
<source>
[PASTE THE REAL NOTES, TRANSCRIPT, DATA OR PREVIOUS DRAFT HERE. Do not summarise it for the model first.]
</source>
<format>
[LENGTH]. [STRUCTURE, e.g. three short paragraphs, no bullet points, subject line first]. [TONE IN CONCRETE TERMS, e.g. direct and factual, the way a project lead writes to a peer].
</format>
<do_not>
Do not use [BANNED PHRASES]. Do not invent [FACTS YOU CANNOT VERIFY]. Do not [KNOWN FAILURE MODE OF THIS MODEL ON THIS TASK].
If anything in <source> is missing or contradictory, list it under “Assumptions” at the end instead of guessing.
</do_not>
That last line is doing quiet, heavy work. It converts silent hallucination into a visible list you can check in ten seconds, which is the difference between a draft you can ship and a draft you have to re-verify line by line.
How do you apply this to a real task at work?
Take a task you repeat weekly and fill the template once with real content, then reuse it. The gain is not one better output. It is that the same brief produces comparable output every week, which is what “consistent” actually means in practice.
Take a recurring example: turning a 45-minute client call into a follow-up email. The persona version says “you are an experienced account manager, write a follow-up email” and produces something polite, padded and slightly wrong about the deadline.
The context version names the artefact as a 250-word email whose purpose is confirming three decisions, names the reader as the client's operations lead who missed the last call, pastes the raw transcript, fixes the format at three paragraphs with a bulleted action list, and forbids inventing dates that do not appear in the transcript.
The second prompt is longer to write the first time and faster every time after, because only the pasted source changes. That is the trade the technique asks you to make.
Once a template stops changing, stop retyping it. Saving a stable brief as a reusable file is exactly what the Agent Skills format was designed for, and the mechanics are covered in our guide to writing your first SKILL.md file.
When is a role instruction still worth including?
Keep the role when it changes vocabulary, audience or omissions. Drop it when the task has one correct answer. That single test decides it for almost every prompt, and it is a much better rule than either “always use a persona” or “personas are dead”.
Keep the role for these
--- Rewriting the same explanation for a different reader, where “explain this to a finance director” and “explain this to a new hire” genuinely produce different, correct answers.
--- Voice and character work in creative drafting, where the persona is the point rather than a performance enhancer.
--- Adversarial review, where “read this as the most sceptical person in the room” changes what gets flagged.
Drop the role for these
--- Extraction, classification and tagging, where the output is determined by the source.
--- Factual question-answering, where a confident persona can make a wrong answer sound more authoritative rather than more correct.
--- Any prompt already running long, where the persona is competing for context window space with your actual source material.
One honest limitation applies to all of this. These are patterns that hold across current mainstream models, not audited benchmark results, and model behaviour shifts with each release. Treat the rule as directional and verify it on your own recurring task rather than trusting any prompting article, including this one.
How do you test this on your own work in ten minutes?
Run a two-version comparison on a task you already know the right answer to. Ten minutes is enough, and using a task with a known good answer is what makes the result readable rather than a matter of taste.
The ten-minute test
--- Minute 1 to 2. Pick one output you produced manually in the last week and were happy with. That is your benchmark.
--- Minute 3 to 4. Write the persona version: “You are an experienced [ROLE]. Write [THING].” Run it. Save the output.
--- Minute 5 to 8. Fill in the five-block template above with the same underlying material. Run it in the same model, in a fresh conversation.
--- Minute 9 to 10. Score both against your benchmark on three questions only: how many facts are wrong, how many sentences would you delete, and how long would each take to fix.
Run the same test in a second model before you conclude anything about the technique, because a difference that appears in one model and vanishes in another is a model quirk rather than a prompting principle.
Keep whichever prompt won and delete the other. The point of the exercise is not to prove an article right. It is to end the guessing about why your outputs vary, which is the actual problem.
What should you change tomorrow morning?
Delete the persona line from your three most-used prompts and replace it with a source block and a negative constraint. That single edit addresses the two most common causes of inconsistent output: missing facts and unrepeated failure modes.
The reframe is small but it changes the job. You are no longer hunting for the sentence that unlocks the model. You are assembling the context a capable colleague would need, and there is no wording that substitutes for material you did not supply.
That is also the honest reason prompting felt like a dark art for three years. It was never mystical. It was under-specified, and the fix is unglamorous and repeatable.
We know AI's cold edges. We know your real challenges. 28 years with UD, turning technology into a partnership with warmth.
Reviewed by the UD AI team.
Find out where your AI skills actually sit
Most people using AI every day have never measured how much of it they are using. UD's AI IQ Test gives you a score and a plain-language read on where your gaps are, in a few minutes and at no cost. If you would rather turn that into a working system, our team will walk you through every step, from prompt templates to a workflow that runs the same way every week.
