A dissection. Four ingredients in the prompt, six classes of model behind it, and a way to figure out which one fits your job.
Who is writing.
The model collapses to whoever you say it is. A vague prompt asks 'an AI' to do the work — you get an AI's median answer. Hand it a job title with judgment and it stops auditioning.
“you are a helpful AI assistant”
You are a careful editor with a bias toward plain language.
why · Specifics inherit specifics. Generic role → generic output.
The verb.
One sentence. One verb. If you cannot say it in plain English, the model cannot do it at all. The task line is where most prompts go to die — buried under hedges and qualifications.
“can you maybe help me with my email and like, make it better”
Rewrite the email below for clarity and a calmer tone.
why · A real verb (rewrite, summarize, extract, draft) gives the model an action. 'Help' gives it nothing.
Audience and constraints.
Who is reading. How long. What format. The boring details that decide whether it lands. Without context, the model defaults to the average reader of the average blog — which is rarely your audience.
“make it good and professional”
Audience: a board chair. Length: 3–4 sentences. Format: plain text, no markdown.
why · 'Professional' is a vibe. 'Board chair, 3 sentences, no markdown' is a target.
Guardrails and tells.
What to do, what not to do, what counts as failure. Rules are how you say no without arguing later. Be specific — 'no clichés' is hand-waving; 'no AI clichés (delve, leverage, in today's fast-paced)' is operable.
“don't be cringe”
Lead with the most important point. No hedges. Banned words: delve, leverage, robust, in today's fast-paced.
why · Models love rules. They especially love rules they can check themselves against.
Read your prompt out loud. Ask each question below. If any answer feels squishy, you found the bone that's about to break in production.
Most failures are visible before you ever hit run. You just have to look.
Could a junior on the team read the role line and act on it?
FIX Replace 'expert' with a job title plus a stance. 'Editor with a bias toward plain language' beats 'world-class writing expert' every time.
Is there exactly one action verb in the task line?
FIX If you have two, split the prompt into two prompts. If you have zero, you wrote a wish, not a task.
Could the model produce a wrong-but-plausible answer for the wrong audience?
FIX Name the reader. 'For a clinician' and 'for a 14-year-old' produce wildly different paragraphs from the same task.
Did you say what shape the output should be?
FIX JSON schema, sentence count, max words, bullet vs prose. The model will guess if you don't tell it, and it will guess wrong on the high-stakes one.
Did you tell it what 'bad' looks like?
FIX Banned words, banned shapes, banned moves. 'No filler. No clichés. No "as an AI" disclaimers.' Three lines saves three rounds.
Could you read the prompt to a smart human who's never seen the task and have them do it?
FIX If they'd ask follow-up questions, the model is silently making those answers up.
'You are a world-class, award-winning, deeply experienced…'
Confusing motivation with instruction. The model has no morale.
Cut the adjectives. Keep the role. Add the stance.
Twelve rules, six examples, three personas, and a poem.
Hedging against failure by piling on. The signal drowns.
Pick the four rules that matter. Move the rest to a fallback prompt.
'Make it good. Professional. But also fun. Not too long.'
Using adjectives where measurements belong.
Replace each adjective with a number, an example, or a banned alternative.
The actual task appears in paragraph four, after the backstory.
Writing the prompt like a memo instead of a spec.
Task line first. Context second. Rules third. Input last.
No format, no length, no failure conditions.
Trusting the model to read your mind. It can't.
Close the door. State the shape. State what 'wrong' looks like.
A perfect prompt aimed at the wrong model still loses. These are the six classes you actually need to know — what they cost, where they shine, and the moment they quietly start lying to you.
The line cook. Volume work, instant answers.
small · 4–8B class · sub-second
best for
not for
If the task fits on a Post-it and the right answer is obvious to a human, this tier is the answer.
The senior IC. Most production traffic lives here.
mid · general purpose
best for
not for
Default here. Step down to fast when latency or cost is the constraint. Step up to heavy only when you can prove the workhorse failed.
The principal. Big context, hard problems, infrequent calls.
flagship · 200k–1M context
best for
not for
Use it as the editor, not the writer. Workhorse drafts, frontier polishes. Cost goes down, quality goes up.
The mathematician. Thinks before it speaks. Literally.
o-series · thinking models
best for
not for
Hand it problems where a human would pull out a notepad. Hand it chat and it'll over-engineer a hello.
The pair of eyes. Reads screenshots, charts, PDFs, the wild.
multimodal · image + video in
best for
not for
If the answer is in the image, use vision. If the image is decoration, you're paying twice for nothing.
The librarian. Not a chatbot. A coordinate system.
vector · semantic similarity
best for
not for
If you said 'the model should know about my data,' the answer is almost always embeddings + retrieval, not a bigger context window.
You almost never need one big model doing one big call. The teams shipping the best work pipe a cheap model into a smart one — or a smart one into a frontier one. Four patterns cover most jobs.
Fast (classify) → Balanced (do the job)
Triage cheap, execute well. Stops the workhorse from wasting cycles on stuff a 4B model could route.
Balanced (draft) → Heavy (critique & polish)
Most of the words come from the cheaper model. The expensive model only touches the final 10%.
Embeddings (find context) → Balanced or Heavy (answer)
Don't stuff the prompt with everything. Search first, then hand the model the five paragraphs that matter.
Reasoning (decompose) → Balanced (do each step)
Reasoning models are great at planning, expensive at typing. Let them outline, let the workhorse fill in.
Describe the job — the volume, the latency, the stakes, what the input looks like — and the advisor recommends a tier (and a stack pattern, if it earns it). Editorial, not enthusiastic. It will push back.
Not a sales pitch. A second opinion before you wire up the wrong thing.
live · the advisor
Describe what you're actually trying to ship — volume, latency, stakes, what kind of input — and the advisor will steer you to a tier (and a stack, if it helps). Or pick one of these to start:
Meerkat asks the four questions for you and assembles the prompt in the format the model actually understands. Sixty seconds, no incantations.