What Is an AI Spec, and Why Did Prompting Tricks Stop Working?
An AI spec is a written brief you hand the model instead of a clever prompt. It states the goal, the audience, the output format, the evaluation criteria, and what to do when the model is uncertain. It is a job description, not an incantation.
The reason it beats tricks is unglamorous. Reasoning models already do internally what the old tricks tried to force from outside.
Wharton's Generative AI Labs has been testing this directly. Their Prompting Science Report 1 found that popular prompt formatting tricks produced gains that vanished once results were aggregated across a full question set. The effect was real on individual items and noise at scale.
Ethan Mollick, who co-authors that series, put the practical version bluntly in July 2026: prompting tricks stopped being very valuable, and the skill that matters now is specifying goals, outputs, evaluation criteria and testing methods. That is management, not wordsmithing.
If you have been collecting prompt phrases for two years, this is good news. The thing that replaced them is learnable in an afternoon and reusable forever.
Do Expert Personas and "Think Step by Step" Actually Improve Output?
Largely no, on current models. Wharton's Prompting Science Report 4 tested expert personas against a no-persona baseline on GPQA Diamond and MMLU-Pro across six models and found no consistent accuracy benefit. Mismatched personas sometimes made results worse. Low-knowledge personas, such as asking the model to answer as a child, reliably reduced accuracy.
The "let's think step by step" instruction has a similar story. On models with built-in reasoning, it adds latency without a matching quality gain, because the model was already going to reason.
This does not mean role framing is useless. It means role framing is not where the leverage is.
What the persona was secretly doing was smuggling in context. "You are a senior B2B copywriter" implies an audience, a register, a length, and a set of conventions. When you write those out explicitly, you get the benefit without depending on the model to infer it.
So the practical move is not to delete "you are a marketing expert" and hope. It is to replace it with the four or five things that phrase was standing in for.
What Are the Five Parts of a Working AI Spec?
A usable spec has five parts: the job to be done, the audience and context, the output contract, the evaluation criteria, and the uncertainty rule. Miss any one and the model fills the gap with an average guess, which is exactly where inconsistent output comes from.
1. The job
--- One sentence stating the outcome, not the activity. "Produce a 300-word LinkedIn post that gets HR managers to download the salary benchmark report" beats "write a LinkedIn post about the report".
2. Audience and context
--- Who reads it, what they already know, what they are sceptical about, and any facts the model cannot invent. This is where you paste real source material rather than trusting recall.
3. The output contract
--- Length, structure, format, and what must not appear. Bans matter as much as requirements: no bullet points, no rhetorical questions, no em dashes, no closing summary paragraph.
4. Evaluation criteria
--- The three to five tests a good answer passes. This is the part almost everyone skips, and it is the part that makes output consistent. Write the rubric you would use to grade a junior's draft.
5. The uncertainty rule
--- What to do when the model does not know. "If a figure is not in the source material I pasted, write [NEEDS SOURCE] rather than estimating" converts a hallucination risk into a visible to-do item.
What Does a Complete AI Spec Look Like?
Below is a full spec you can paste into ChatGPT, Claude, or Gemini and adapt in about two minutes. It is deliberately mundane. Specs work because they are boring and complete, not because they are clever.
Copy this, replace the bracketed parts, and paste your source material at the bottom.
Try this spec:
JOB
Produce [one 900-word article] whose purpose is [getting operations managers in Hong Kong to book a process review]. Success means [a reader forwards it to their manager].
AUDIENCE
Readers are [operations managers at 20 to 200 person HK companies]. They already know [what workflow automation is]. They are sceptical about [vendor claims of 10x savings]. They do not know [how to scope a first project].
OUTPUT CONTRACT
--- Length: 850 to 950 words
--- Structure: 5 sections, each with a question as its heading
--- Each section opens with a 2-sentence direct answer
--- Format: plain paragraphs, maximum 3 lines each
--- Do not use: bullet lists, em dashes, rhetorical questions, the words "leverage", "unlock", "landscape"
--- Do not end with a summary paragraph
EVALUATION CRITERIA
A good draft passes all of these:
--- Every claim has a number, a named source, or a concrete scenario attached
--- A reader could act on section 3 today without asking me a question
--- No sentence would embarrass me if a client quoted it back
--- Removing any paragraph would lose real information
UNCERTAINTY RULE
If a statistic, price, or date is not in the source material below, write [NEEDS SOURCE] instead of estimating. Do not infer figures from adjacent facts.
PROCESS
Before writing, list the 5 section headings and wait for my approval. After I approve, write the draft, then score it against the evaluation criteria and tell me which criterion is weakest.
SOURCE MATERIAL
[paste your notes, transcript, data, or existing copy here]
The last two blocks are what most people leave out. The process block turns one long guess into two short checkpoints. The self-scoring instruction is unusually effective, because the model has to re-read its own draft against your rubric rather than yours.
How Do You Reuse One Spec Across a Month of Work?
Save the spec as a file, not as a chat message. The audience block, output contract, and evaluation criteria stay identical across every piece of the same type. Only the job line and source material change, which turns a 15-minute prompt-wrangling session into a 90-second paste.
In practice that means keeping one spec per output type. A weekly client update spec. A cold outreach email spec. A meeting-notes-to-actions spec. Three files cover most of a marketer's or ops manager's recurring output.
Then put the stable part somewhere the model reads automatically. ChatGPT Projects, Claude Projects, and Gemini Gems all accept persistent instructions attached to a workspace. Paste the audience block, output contract and evaluation criteria there once, and every new chat in that workspace starts pre-loaded.
The version-control habit matters more than the tool. When a draft comes back wrong, do not correct the draft. Add the missing rule to the spec file, then rerun. After roughly six rounds of that, the spec stops producing surprises, and the improvement is permanent instead of trapped in one conversation.
This also keeps your instructions out of the part of the conversation that degrades. Long threads lose fidelity to their earliest instructions, a failure mode covered in this breakdown of context rot. A spec that lives in project instructions or gets repasted at the top of a fresh chat sidesteps the problem entirely.
Where Do AI Specs Break Down?
Specs fail in four predictable ways: they get too long, the criteria contradict each other, they over-constrain creative work, and they are treated as write-once documents. All four are fixable, and knowing them in advance saves a week of confusion.
Over-length
--- Past roughly 600 words of instruction, models start weighting your rules unevenly. If your spec is growing past that, you are probably specifying two different output types and should split it into two files.
Contradictory criteria
--- "Be comprehensive" plus "maximum 400 words" is a conflict, and the model resolves it silently by picking one. Read your criteria as a set and delete whichever rule you would sacrifice in a real deadline.
Over-constraining creative work
--- Specs are excellent for repeatable output and actively harmful for ideation. When you want twenty strange angles, a tight output contract gives you twenty variations of the same safe angle. Run divergent work with a loose brief and switch to the spec only at drafting.
Treating it as finished
--- A spec written once and never revised drifts out of date as your product, audience, and the model itself change. Reasoning depth in particular is now adjustable on several models, which changes what a spec needs to say about effort. UD's earlier piece on adaptive thinking covers how that dial interacts with instruction-following.
How Can You Test This in the Next Twenty Minutes?
Pick the one output you produce most often, then run the same task twice: once with your usual prompt, once with the five-part spec above. Compare the two drafts against your own evaluation criteria rather than against your gut. The gap is usually obvious on the first attempt.
Do it on something real, not a test topic. Your instincts about whether output is usable are far sharper on work you actually have to ship.
Two things tend to happen. The spec draft needs less editing, which is the point. And writing the evaluation criteria forces you to articulate a standard you had been carrying around implicitly, which improves your own briefs to human colleagues as a side effect.
That is the honest summary of where prompting went. The clever phrasing era is over, and what replaced it is the ordinary discipline of saying clearly what good looks like before you ask for it. It is less fun and considerably more reliable.
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.
Know Where Your AI Skills Actually Stand
Writing a spec is one level of AI skill. There are several above it, and most people cannot tell which one they are on.
UD's free AI IQ Test scores your working knowledge in 15 questions, then shows you the specific gaps worth closing next. We'll walk you through every step, from reading your result to building the workflow that fixes the weakest link.
