Guides
The checks explain the nine things that break a prompt. The techniques cover how to do the harder parts well. Every guide shows a real prompt and runs the real checker on it — the checks in these articles are live, not screenshots.
The nine checks
One guide per check: the failure it catches, why it costs you, and the fix.
Why “YOU MUST” makes your prompt worse
Forceful phrasing feels like it should force compliance. On current models it backfires — they follow instructions literally, so shouting makes them overtrigger.
Read the guide →Separate your data from your instructions
Paste a long document straight into a prompt and the model has to guess where your instructions end. Anything inside that looks like a command may get followed.
Read the guide →State a concrete task, not a topic
A topic is not an instruction. Without a verb naming the job, the model has to guess whether you want a summary, an essay, or a list.
Read the guide →“Make it good” tells the model nothing
Good, better, engaging, appropriate — these are judgements, not instructions. The model cannot check its work against a word that means something different to everyone.
Read the guide →Say what shape the answer should take
Leave the format unsaid and models default to prose you then have to parse by hand. One sentence naming the format saves the round trip.
Read the guide →The model cannot see what you can see
The same question answered for a beginner and for a specialist are different answers. If you do not say who it is for, the model picks for you.
Read the guide →“Fix it” — fix what?
A pronoun with nothing to point at is a coin flip. The model cannot see the tab you are looking at, the file you have open, or the thing you meant.
Read the guide →One example beats a paragraph of description
Describing a format in prose is slow and lossy. One worked example pins down structure, tone, and edge cases in a way instructions rarely manage.
Read the guide →Tell the model how to know it succeeded
Without a bar to clear, the model cannot check its own work. Naming what to include and what to leave out gives it something to test the answer against.
Read the guide →
Techniques
The methods behind good prompts — when each one earns its place, and when it is just superstition.
Few-shot prompting: when examples beat instructions
A worked example pins down structure and edge cases that a paragraph of description cannot. Here is how many to give, and when one is already too many.
Read the guide →Role prompting: when “act as a…” actually helps
Assigning a role is half technique, half superstition. It works when it carries real constraints the model would otherwise miss — and does nothing when it is just flattery.
Read the guide →Reasoning prompts: making the model show its work
“Think step by step” is folklore that sometimes helps and sometimes wastes tokens. The honest version: what it does, when reasoning models make it redundant, and how to ask.
Read the guide →Iterating on a prompt: fixing the part that is actually wrong
When the first output disappoints, the instinct is to rewrite everything. The faster route is to read the failure, change one thing, and know which one.
Read the guide →