// Contents
The phrase “magic words for a neural network” lasts until the first serious task. Then it turns out the result depends not on incantations but on clarity: what is needed, for whom, in what form, and how you will know it turned out well. That is closer to a technical brief than a spell.
Five parts of a request
- Task. What exactly to do.
- Context. For whom, why, which facts are known.
- Format. What the answer should look like: list, table, JSON, length.
- Constraints. What to avoid, the tone, banned words.
- Quality criteria. How you will judge the result.
A weak request:
Write a post about neural networks.A stronger one:
Task: write a Telegram channel post on how small businesses use AI.
Context: audience is cafe and shop owners with no technical background.
Format: up to 900 characters, a hook in the first line, one concrete example.
Constraints: no "revolution" or "the future is here", no emoji.
Criterion: the example must be doable in one evening.What works for text models
- Examples. One or two samples of the desired result beat a long description of style.
- Structured output. If the answer goes into a program, ask for JSON with a defined schema and validate it with code, not by eye.
- Step breakdown. For complex tasks ask first for a plan or a list of assumptions, then the result itself.
- Self-checking. Ask separately for a list of claims that need source verification.
- A role (“you are an experienced editor”) helps less than concrete criteria.
Image models
An image is described like a shot: subject, setting, light, angle, style, color palette. For reproducibility keep part of the style description fixed in every request. State what must not appear (“no text, no logos”). English sometimes gives more predictable results on visual models, but modern models understand other languages well too: test on your model.
Video
Video models are sensitive to motion. Describe the action and the camera separately: what happens, how the camera moves, how many seconds. One shot, one action. If you need a stable character, start from a reference image rather than text.
Music
State genre, tempo, mood, instruments and structure (“8-bar intro, verse, no vocals”). Mind the rights and terms of use for generated tracks on your platform.
Iterate instead of aiming for the perfect prompt
Do not try to write the perfect prompt first time. Run it, see what exactly is wrong, and change one thing at a time. Keep good prompts in a file: a prompt library pays off quickly.
How to check the result
For important tasks, build a small set of test examples and run every new prompt version against it. It is the same principle as in software: a change should improve things, not just look better on the last example.
FAQ
What is an AI prompt?
An AI prompt is an instruction given to a model. It usually includes the task, context, desired format, constraints, and examples or quality criteria.
Why do AI prompts fail?
Most prompts fail because they are too vague. The model needs audience, format, style, constraints, and a clear definition of what a good result means.
Should prompts be written in English?
For many image and video models, English can be more predictable. For local-language writing, it is often better to prompt in the target language and define the tone clearly.
