10 habits for writing better prompts
None of these need special tricks. They're the habits we use when we write the prompts in this library, and each one links to a free prompt that shows it.
1. Say what you want and why
"Summarise this" gets you a generic summary. "Summarise this so I can decide whether to renew the contract" gets you the parts that matter. Give the goal and who the result is for.
See it in: Summarise a long document for decisions
2. Give it the context it can't guess
The model only knows what you paste in. Include what you expected, what actually happened, and the real error, code or text. Missing context is the most common reason for a vague answer.
See it in: Root-cause a bug, don't patch symptoms
3. Ask for a specific output format
Say exactly what the answer should look like: "5 bullet points", "a table with these columns", "JSON with these keys". It's easier to read, and easier to paste into whatever comes next.
See it in: Small-model friendly summariser
4. Say what not to do
Models like to be helpful, which can mean adding ideas, rewriting too much or changing things you didn't ask about. A short "don't" list keeps the edit focused.
See it in: Edit for clarity, keep my voice
5. Give it permission to say "I don't know"
Tell the model to say "unverifiable", mark a TODO, or ask a question instead of guessing. It won't stop every made-up fact, but it makes guesses easier to spot.
See it in: Fact-check this claim
6. Show an example
If you want a particular style or structure, two or three examples usually work better than describing it. This is especially true for small local models.
See it in: Few-shot style mimic (small models)
7. Break big jobs into steps
Ask for a plan first, then one step at a time, and check each step before the next. For coding agents, this means fewer runaway edits across files you didn't mean to touch.
See it in: One small step at a time
8. Keep your instructions separate from your material
Put pasted text inside clear markers such as """ or a code block, and keep the instructions outside them. That way the model doesn't confuse a sentence in your document with an instruction.
See it in: Structured JSON extraction (small models)
9. Match the prompt to the model
A long, nuanced prompt that works on a frontier model can confuse a 3B local model. For small models, use shorter prompts with fixed choices and strict formats. Each prompt's model notes tell you what it was written for.
See it in: Classify into fixed labels
10. Save what works and reuse it
When a prompt gives you a good result, save it and swap the specifics for {{variables}}. Next time you only fill in the blanks. That's how every prompt in this library is built. For tasks you repeat a lot, turn the prompt into standing instructions for an agent.
See it in: Turn a repeated task into agent instructions
Whatever habits you use, always review AI output before you rely on it.