Most prompt engineering advice is folklore passed between screenshots. Some of it works, a lot of it is superstition, and almost none of it explains why. This guide covers the six things that reliably change what Claude produces, and it ends with the case for stopping prompt engineering altogether once you are doing the same task repeatedly.
1. Say who is asking and what happens next
The single biggest lever is context about the situation, not clever phrasing. "Write a product description" and "Write a product description for a 200-euro ceramic kettle sold to design-conscious buyers on our own store, to sit under three photographs, about 60 words, no exclamation marks" are not the same request. The second one has an audience, a channel, a length and a constraint. The model was never going to guess any of those.
The habit to build: before you type the instruction, answer four things in your head and then put them in the prompt. Who reads this. Where does it appear. How long. What must it not do.
2. Give examples, not adjectives
"Make it punchy" means nothing consistent. Two sentences you consider punchy mean something exact. Paste one good example and one bad example and say which is which, and the output moves further than any amount of describing the tone.
This is the technique with the best effort-to-result ratio in the whole list, and it is the one people skip because finding an example takes thirty seconds and typing "punchy" takes one.
3. Ask for the reasoning before the answer, when the task is analytical
For anything involving comparison, diagnosis or numbers, ask Claude to work through it before committing to a conclusion. "Go through each clause and note what it does before you tell me whether to sign" produces different quality than "should I sign this".
The caveat: this helps on analytical work and does nothing for creative or formatting tasks. Applying it everywhere just makes the answer longer.
4. Specify the shape of the output
If you want a table, ask for a table and name the columns. If you want it to fit in a Slack message, say so. If you want no preamble, say "start directly with the first item, no introduction". Claude defaults to a helpful essay because that is the safe general answer, and the safe general answer is rarely the one you wanted.
5. Tell it what to do when it does not know
Unspecified, the default behaviour on a gap is to produce something plausible. Specified, it is not. "If the document does not say, write UNSTATED rather than inferring" costs eight words and removes most invented detail from analytical work.
This is the technique that matters most on legal, financial and medical documents, and it is almost never in the popular guides.
6. Iterate on the instruction, not the output
When the answer is wrong, most people fix the answer in a follow-up message. That fixes one answer. The alternative is to work out what was missing from the instruction and add it, so the next twenty answers are right too.
Concretely: instead of "make it shorter", ask yourself what length you actually wanted, and put that number in the original prompt. Instead of "less formal", find the sentence that was too formal and show it what you would have written.
The part where prompt engineering stops being worth it
Everything above is worth learning. But notice what happens when you get good at it: your prompts get long, they encode a lot of decisions, and you retype them constantly.
At that point you are not writing a prompt any more. You are maintaining a specification, badly, in a chat box that forgets it. The fix is to move that specification out of the message and into Custom Instructions, where it persists across every conversation until you change it. That is all a skill file is: a well-tested specification for one kind of work, written once.
The rule of thumb is simple. Doing this task once, write a good prompt. Doing it weekly, write it down. Doing it daily, use a file someone has already tested against a hundred variations of the task.
What a skill file looks like
It is plain text. A role, how that role works, what it always asks before answering, what output format it produces, and what it refuses to guess at. You paste it into Custom Instructions once and every conversation starts from there. Setup is under five minutes and there is a step-by-step guide here.
Try a free one before buying anything, or browse the catalogue by role.
KissMySkills is an independent marketplace and is not affiliated with, endorsed by, or connected to Anthropic, OpenAI, or any other AI company.