The short answer
Prompt engineering is writing instructions that get reliable, useful output from AI models. It matters more than ever, but it is not magic words: it is clear thinking written down, starting with a definition of success, then giving context, structure and examples, then testing and refining.
What it is
A prompt is everything you tell a model before it answers: role, context, task, constraints, output format and examples. There are three families of prompts in U3’s work: language model prompts for writing, analysis and code; image prompts; and video prompts. Each family has its own shape, but the core is the same: clear intent, enough context, explicit constraints, a stable structure and a way to verify.
Anthropic’s prompting best practices start from general principles: be clear and direct, add context and the motivation behind instructions, use examples, structure prompts with XML tags, and give the model a role. Its golden rule: show your prompt to a colleague with minimal context on the task; if they would be confused, the model will be too.[1]
The same documentation asks for three things before any prompt engineering: a clear definition of success criteria, a way to test against them empirically, and a first draft prompt.[2]
The prompt as a creative brief
In the U3 Method a good prompt is a complete brief in seven parts:
- Role: who is speaking and who they write for.
- Context: what the model must know about the project.
- Goal: what the output must achieve.
- Constraints: length, tone, language and what is forbidden.
- Format: the structure of the answer.
- Example: a sample of what you consider success.
- Check: how you will judge the result.
Prompts by medium
The same core, different shapes:
- Text: role, context, task, output format and an example.
- Image: eleven layer anatomy and ranked references.
- Video: a timecoded prompt from 0:00, tags for elements and one action per range.
- Code: a requirements document, then a plan reviewed by a person, then execution in small steps.
For images specifically, Google recommends positive framing: describe what you want instead of what you do not want, “empty street” rather than “no cars”.[3]
Testing a prompt
U3’s refinement loop:
- Write a success criterion that can be judged.
- Write the first version as a complete brief.
- Run it more than once, because one result proves nothing.
- Change one variable per round.
- Save the winning version as a dated template in the bible.
Example
Weak prompt: “write a description for a perfume”. Prompt as a brief: “You are a copywriter for a luxury fragrance brand in the Gulf. Write a description of a warm woody perfume for a product page in an online store. Audience: men aged 30 to 45. Tone: calm and confident, never exaggerated. Length: 60 words in modern standard Arabic. Do not mention prices or promise results. Return three versions with different openings.” The difference is not length; it is the number of decisions.
In U3 practice
At U3 prompts do not live in scattered chats. Successful templates are saved in the project bible with their date and the engine they were tested on, and prompts are written from the storyboard and shot list, not from memory. That way a prompt becomes a reusable asset, not an attempt that gets lost. When a new model is released, the saved templates are tested on it first, so we learn quickly what still works and what needs adjusting.
Arabic in prompts
Arabic is not one language in daily use. If you ask for text “in Arabic” without specifying, you will usually get a generic standard Arabic that does not sound like your audience. Name the dialect explicitly: modern Egyptian for a social ad in Cairo, Gulf Arabic for an audience in Riyadh and Dubai, or clear modern standard Arabic for formal or legal text. And give a short example in the tone you want, because an example teaches the model more than a description.
Prompts that run inside products
When a prompt becomes part of a product, such as a support assistant, a summarizer or a message classifier, the success criterion changes. The question is no longer “is the result good?” but “is it good every time, for every user and every odd input?”. That is why system prompts are written with stable, clear instructions, a structured output format the code can read, examples of the hard cases, and a set of tests run after every change.
Before a prompt goes into a product:
- Try it on real user inputs, not ideal examples.
- Try it on empty, vague or other language inputs.
- Decide what happens when the model does not know the answer.
- Record the version and date so you know which change altered behaviour.
Where to start
In AI From Zero the prompt is taught as a creative brief from day one. Then Image Craft goes deep into image prompts, AI Video Production into timecoded prompts, Claude Code for Real into coding prompts, and AI Inside Products into prompts that run inside systems.
Common mistakes
Mistake: A one line request with no context.
Fix: Write the brief in its seven parts.
Mistake: Hunting for “magic words” instead of decisions.
Fix: Every word needs a decision you can explain behind it.
Mistake: Changing many things at once.
Fix: One variable per round, then compare.
Mistake: Successful prompts get lost in chats.
Fix: Save them as dated templates in the bible.
The course that teaches this
AI From Zero
Your first image in the first lesson, then the full picture: models, tools, prompts and cost.
Frequently asked questions
Sources
Every bracketed number on this page points to one of these sources. All are published, with the date we last checked each one.
- 1Prompting best practices
Anthropic · Accessed 4 October 2026
- 2Prompt engineering overview
Anthropic · Accessed 4 October 2026
- 3Ultimate prompting guide for Nano Banana
Google Cloud · Published 6 March 2026 · Accessed 4 October 2026