When two people get very different answers from the same AI, it's rarely the model's mood. It's that they pulled different levers — and most of those levers aren't even in the prompt.
People treat AI output as if it arrives by luck: ask, and hope. It isn't luck. An answer is the output of a handful of variables you control, plus one or two you don't. Name them and you can move them on purpose. Here are the seven that account for most of the variance, ordered by how much they move the result.
1. The model you chose
This is the single largest lever, and it sits upstream of everything else. A model is the trained system that generates the text; capability varies widely across them. A task that's hard — multi-step reasoning, careful code, weighing trade-offs — can fail on a weaker model no matter how well you phrase the request, and succeed on a stronger one with a sloppy prompt.
Decision rule: if the answer is wrong in a way that looks like the system couldn't, not wouldn't, change the model before you change the prompt. Reaching for better wording on an underpowered model is the most common wasted move.
2. The context you supply
A model has no memory of you, your files, or your earlier conversations unless that information is inside the context — the text it can see for this request. It does not look anything up unless you give it the source. "Summarize my meeting" with no notes attached produces a generic summary of a meeting that never happened.
The fix is mechanical: paste the document, attach the data, include the relevant prior decision. Most "the AI doesn't understand my situation" complaints are really "the AI was never shown my situation."
3. The specificity of the ask
Vague in, vague out. An AI fills unstated requirements with the average of everything it has seen, which is rarely what you wanted. "Write a summary" yields a shapeless paragraph; "Write a 5-bullet summary for a non-technical executive, decisions only" yields the thing you pictured.
Three constraints carry most of the weight: format (bullets, table, length), audience (who reads it), and scope (what to include and exclude). State them and the answer narrows toward your intent.
4. The examples you show it
One or two worked examples often outperform a paragraph of instructions. Showing the model an input paired with the output you'd accept — a technique called few-shot prompting — is usually more reliable than describing that output in words. The model matches the pattern faster than it parses the rules.
Example: rather than explaining your preferred email tone in five sentences, paste one email you'd actually send. The next draft will track it closely.
5. The role and standing instructions
A system instruction is the standing brief — the "be a careful financial analyst who flags assumptions" you set once, before the specific question. It changes the defaults the model reaches for: its vocabulary, how much it hedges, what it assumes you already know.
This lever is invisible because it lives outside the immediate question, which is exactly why it's under-used. Set the role deliberately and you stop re-specifying tone on every single request.
6. The randomness setting
Most tools expose a knob — often called temperature — that controls how much the model varies its wording from one run to the next. High temperature means more variation; low temperature means the same prompt returns close to the same answer each time. This is the lever that explains why an identical question can produce two different answers minutes apart.
Decision rule: for facts, extraction, and code, turn it down — you want repeatable. For brainstorming and first drafts, turn it up — you want range. Reaching for "regenerate" until you like an answer is just pulling this lever blindly; set it on purpose instead.
7. What it cannot see
The last lever is the one you don't control, and naming it saves hours. A model has a knowledge cutoff — a date after which it learned nothing — and no access to your private or real-time data unless you supply it (see lever 2). When an answer is confidently wrong about a recent event or your internal numbers, the model isn't broken. It never had the information.
This distinction is the whole game: "it's wrong" calls for a better prompt; "it can't know that" calls for you to provide the source or accept the limit. Confusing the two leads people to keep rephrasing a question the model was never equipped to answer.
The levers at a glance
| Lever | What it controls | How to pull it |
|---|---|---|
| Model | Ceiling on capability | Upgrade before re-prompting hard tasks |
| Context | What it can see | Attach the actual source |
| Specificity | How close to your intent | State format, audience, scope |
| Examples | Pattern to match | Show one accepted output |
| Role | Default tone and stance | Set a standing instruction |
| Randomness | Repeatability vs. range | Down for facts, up for ideas |
| Knowledge limits | What's impossible | Supply data or accept the gap |
Where to start
Start with lever 1, then lever 2. Most disappointing answers come from a model that's underpowered for the task or starved of context — and both are fixed before you ever touch your wording. The remaining five turn a workable answer into a precise one.
The levers only help if you can find what worked last time. Keep your sharpest prompts, role instructions, and example pairs in one place you can search — a notebook in JustJot.ai turns a lucky answer into a repeatable one.