There is a large industry selling prompt templates, magic phrases and "secret" formulas. Most of it is folklore built on a misunderstanding of what these systems do.
The genuinely effective techniques are unremarkable, and they follow directly from how the models work — see what a large language model actually does.
Three things account for most of the improvement: give it the material instead of asking it to remember; state the format you want; and correct the answer rather than starting again. Role-play openings and magic phrases do far less than the tip industry claims.
1. Supply the material
The single biggest improvement, and the one that follows directly from the mechanism.
These models are strong at transforming text you give them and unreliable at retrieving facts from memory. So move every task you can from the second category into the first.
| Weak | Strong |
|---|---|
| "What does employment law say about notice periods?" | "Here is my contract. What does it say about notice?" |
| "Summarise the article about X" | "Summarise this: [paste]" |
| "What are the main points of that report?" | "Here is the report. Give me five bullet points." |
The left column invites invention. The right column cannot, because the answer is in front of it.
2. Say what you actually want back
Models do not guess your intent, and the default output is generic because "write about X" specifies nothing.
Name whichever of these matter:
- Length — "three sentences", "under 200 words"
- Format — bullets, table, numbered steps, email
- Audience — "for someone with no background in this"
- Tone — plain, formal, technical
- What to leave out — "no introduction", "skip the caveats"
"Explain inflation" produces an encyclopedia entry. "Explain inflation in four sentences to a 15-year-old, no jargon" produces something usable.
3. Correct, don't restart
The instinct on a bad answer is to rewrite the prompt and try again. That throws away everything that was right.
Say what is wrong instead:
- "Too long. Half the length."
- "The third point is wrong — it's the opposite."
- "Keep the structure, make the tone plainer."
Faster, and it converges. Rewriting from scratch often produces a different answer with different problems.
The pattern of typing a prompt, disliking the result and typing a new prompt is the most common way people get poor output. Two or three rounds of correction beat ten fresh attempts, because each round keeps the parts that worked.
4. Ask for the working on anything with steps
For calculations, comparisons and multi-step reasoning, asking the model to show its steps genuinely improves accuracy — the intermediate steps become part of what it is conditioning on.
It also lets you spot where it went wrong instead of only seeing a wrong final number.
5. Give an example when the format matters
If you need output in a specific shape, one example does more than a paragraph of description.
"Rewrite these in this style. Example: 'Q: What is X? A: X is…' Now do these five."
Showing beats describing, especially for anything structured.
What does less than claimed
"You are a world-class expert." Nudges vocabulary and tone slightly. Adds no knowledge whatsoever — the model does not know more because you said it does.
Emotional pressure. "This is very important to me", offers of tips. Widely shared, poorly supported.
Extreme length. A prompt padded with instructions but no material is worse than a short one with the source text attached.
Politeness. Costs nothing and does nothing. Say please if you prefer; it is not a technique.
Things worth knowing about the limits
The context window. There is a maximum amount of text it can consider at once. In a long conversation, early material falls out of range — which is why a chat that started well can drift. Starting fresh with the relevant material pasted in often works better than continuing.
The knowledge cut-off. Training stops at a point in time. Anything after it is unknown unless the tool is actually searching the web, which is a separate feature and not always on.
Confidence is not accuracy. The tone is identical whether it knows or is inventing. Every number, name, date and citation needs checking before you act on it.
A working template
For most non-trivial tasks:
[The material — paste the text, data or document]
[The task — what to do with it]
[The format — length, structure, audience]
[Constraints — what to avoid, what to keep]Not clever, and it removes most of the ambiguity that produces disappointing output.
The rule underneath all of it
The more of the answer you supply, the better the answer you get back.
That sounds like it defeats the purpose. It does not — the value is in the transformation, the structuring, the rewriting and the explaining. Expecting the model to supply the facts and the transformation is where it goes wrong, and it is what almost every prompting complaint comes down to.
Frequently asked questions
Do magic phrases like "you are an expert" help?
Marginally at best, and far less than the prompt-tip industry suggests. Naming a role can nudge tone and vocabulary, but it adds no knowledge. Supplying actual context helps enormously more.
Should prompts be long or short?
Long enough to remove ambiguity, no longer. Length itself does nothing. A short prompt with the source text attached beats a long prompt full of instructions and no material.
Why does it give different answers to the same question?
Because it samples from a probability distribution rather than looking up a stored answer. Some variation is inherent. If you need consistency, supply the source material rather than relying on memory.
What is the single biggest improvement?
Giving it the text instead of asking it to recall the text. Transformation tasks are where these models are strong; retrieval from memory is where they invent things.
Is it better to start over when an answer is wrong?
Usually not. Saying precisely what is wrong is faster and produces better results than rewriting the question, because the model keeps everything that was already right.
Found an error? Email us and we will fix it and note the change at the bottom of this article. Hello@daily-atlas.com
