Most disappointing output from an AI assistant is a briefing failure, not a model failure. The request that produced it would have produced the same shrug from a competent freelancer: no audience, no constraints, no source material, no idea what good looks like. The patterns below are the ones that survive contact with real work. Each has a worked example you can adapt this afternoon.
1. The brief pattern: role, task, audience, constraints, format
The single highest-return change is to stop writing requests and start writing briefs. Five elements: who the model is acting as, what it is producing, who reads it, what limits apply, and what shape the answer takes.
Weak: Write a description for our Photoshop course.
Better: You are writing for a training school's website. Produce a course description for a two-day beginner Photoshop class. Readers are office staff in Colombo who have been told to "sort out the images" and have never opened Photoshop. Constraints: British English, 120 to 150 words, no marketing language, no exclamation marks, name three specific things the learner will produce. Format: two paragraphs, then a bulleted list of three outcomes.
The second prompt is longer, and it is the only one that can be judged. Note that every constraint is checkable: word count, number of outcomes, spelling variety.
2. Ground it in your own material
An assistant answering from memory is guessing about your business. An assistant answering from pasted material is doing comprehension, which it is far better at. Paste the source and say what to do with it.
Below are my notes from a client call, unedited. Using only what is in the notes, produce: (a) a list of agreed actions with owners, (b) a list of open questions where the notes are ambiguous, and (c) anything the client asked for that we have not priced. Do not add anything that is not in the notes. If something is unclear, put it in the open questions list rather than guessing.
The instruction "using only what is in the notes" plus an explicit place to put uncertainty is what keeps invented detail out of the answer. Without a designated home for doubt, the model resolves it silently.
3. Show one example of what good looks like
Describing a tone of voice takes a paragraph and still misses. Showing one example takes ten seconds and lands.
Here is one of our existing course summaries, which the team considers correct in tone and length: "Two days of hands-on Photoshop for beginners: layers, masks, retouching, type and export, with practice mapped to the Adobe Certified Professional exam." Write summaries in exactly that voice and length for the three courses described below.
Two or three examples work better than one when the pattern is subtle. Choose examples that differ from each other, or the model will copy the coincidence rather than the rule.
4. Plan first, write second
For anything longer than a few paragraphs, splitting the job in two produces better work and, more usefully, lets you correct course before a thousand words exist.
Step one only: propose a structure for a 1,200 word article on choosing between Photoshop and Illustrator. Give the section headings and one line on what each covers. Do not write the article yet. Wait for my approval.
Fix the outline, then say "write section three only". Long single-shot documents drift; sections do not.
5. Specify the format as a shape, not an adjective
"Make it structured" is not an instruction. "A table with these four columns" is.
Return a table with the columns: task, application, why, and file format to hand over. One row per task, using these tasks: retouch a portrait, draw a logo, lay out a 12-page brochure, build a social banner. Keep each cell under 15 words.
The same applies to machine-readable output: name the keys you want, state what happens when a value is unknown, and ask for nothing outside the structure.
6. Say what to avoid, and give it permission not to know
Negative constraints work when they are specific. "Don't be generic" achieves nothing; a banned-words list achieves quite a lot.
Avoid these words and constructions entirely: unlock, elevate, seamless, robust, in today's fast-paced world, "whether you are". Do not open any sentence with a participle. If you do not have enough information to write a section accurately, write "NEEDS INPUT" and a one-line question instead of filling it with generalities.
That last clause is the important one. An assistant with no escape route will produce plausible filler, because filler satisfies the request. Give it somewhere to put its uncertainty and you can see it.
7. Ask for options that differ on purpose
Requesting "five headlines" gets five phrasings of the same idea. Requesting five different angles gets five ideas.
Give five subject lines for an email announcing an evening Illustrator class. Each must take a different angle: one about the outcome, one about the time commitment, one about a specific mistake beginners make, one about certification, one plain and factual. Maximum 55 characters each. Label the angle before each line.
8. Critique, then revise
Models are better at finding faults in text than at avoiding them while writing. Use that as a second pass.
Here is the draft. Identify the three weakest sentences and say precisely what is wrong with each: vague, unsupported, repeats an earlier point, or marketing filler. Then rewrite only those three. Leave everything else exactly as it is.
"Leave everything else exactly as it is" matters. Without it you get a full rewrite and lose the parts you liked.
9. Give it the marking scheme
If you know how you will judge the output, say so before it is produced, and ask for a check afterwards.
The finished text will be judged on four things: every claim is checkable against the course outline below, the reading age is suitable for a non-designer, it is under 300 words, and it contains no invented numbers. Write it, then list the four criteria and state pass or fail for each with a one-line justification.
Self-assessment is not proof, but it surfaces obvious misses and it costs one line.
10. Make it interview you
Useful when you know what you want but cannot articulate it.
I need a one-page proposal for a corporate training client. Before writing anything, ask me up to eight questions that would materially change the proposal. Ask them one at a time and wait for my answer.
What does not work
Politeness makes no measurable difference to quality, though it costs nothing. Grand role framing such as "you are the world's greatest copywriter" adds far less than a concrete audience description. Threats and offers of payment are folklore. Very long prompts that bury the actual instruction in the middle perform worse than short ones with the instruction at the end. Asking for facts, statistics or citations without supplying source material invites confident invention, so supply the sources or verify every number yourself. And asking a model to count characters, words or occurrences precisely is asking it to do the one thing it is structurally bad at, so check those by hand.
Keep the ones that work
The teams that get real value are not the ones with clever prompts. They are the ones that saved the prompt that worked, put it in a shared document with a note on when to use it, and stopped rewriting it from scratch every Monday. Start that file today, even if it has three entries in it.
Our prompt engineering course builds exactly that library, applied to your own work, with a session on evaluating output so you can tell whether a change to a prompt actually improved anything.
