"I Tried It and the Output Was Rubbish"
This is the most common thing people say after their first serious attempt at using AI for work, and it is almost always true. The output was rubbish. What they usually have not noticed is what they asked for: eleven words, no context, no example, no statement of who it was for. The model did a reasonable job of an impossible request.
There is a whole industry around "prompt engineering" that makes this sound more mystical than it is. There are no magic words. There is no phrasing that unlocks a secret better model. What there is: a short list of things a good instruction contains, most of which you would include automatically if you were briefing a competent freelancer who had never met your business.
The Five Parts
The task, stated as a verb
"Draft", "summarise", "extract", "rewrite", "compare". Vague openers like "help me with" or "look at this" leave the model guessing at what kind of output you want, and it will guess wrong in an expensive-to-notice way.
Who it is for, and what they already know
"For a client who has not seen the project before" and "for our site foreman who knows the job" produce completely different documents. This one line removes most of the mismatch between what you got and what you wanted.
The material it needs
The model knows nothing about your business. Paste the thread, the notes, the previous version, the numbers. Most bad output is not a reasoning failure — it is the model filling a gap you left, which it will always do rather than asking.
An example of good
The highest-leverage line in this article. Attach your best previous version of the same document and say "match this style". This changes results more than model choice, more than clever phrasing, more than anything else on this list.
The constraints that matter
Length, format, tone, and — importantly — what not to do. "Do not invent figures. If something is missing, list it as a gap instead of filling it." That one instruction prevents a whole category of failure.
Why the Example Does So Much Work
Adjectives are ambiguous and examples are not. "Professional" means one thing in a law firm and something quite different on a construction site. "Concise" could mean four sentences or four paragraphs. "Friendly" is anybody's guess. You can spend three rounds of iteration converging on a register that one attached document would have conveyed instantly.
It also solves a problem you may not have articulated: your standard is mostly tacit. You know a good quote when you see one, but writing down the rules that make it good is genuinely hard. You do not have to. Show the model three good ones and the pattern comes across, including the parts you could not have named.
A Thin Prompt and a Real One
Write a follow-up email for a client who has not responded to our quote.Draft a follow-up email to a client who has not responded to a quote sent
eleven days ago.
Context — the quote and the original thread:
[paste both]
Audience: the owner of a mid-sized construction firm. Busy, direct, has
dealt with us twice before. Not a technical reader.
Match the style of this email I sent last month, which worked well:
[paste your best previous follow-up]
Constraints:
- Under 120 words
- No apology for following up, and no "just checking in"
- Reference one specific detail from their project
- One clear next step
- Do not invent any figures. If a number is missing, mark it [FIGURE]
rather than filling it in.The second one takes four minutes the first time. The difference is that the second one is reusable: everything except the pasted material stays the same next week, which is what turns a good result into something that runs.
The Habits That Help Most
Works
- Attaching an example of good output
- Saying who the reader is and what they know
- Asking it to list gaps rather than fill them
- Starting a fresh conversation for a new task
- Asking for a shorter version — it is nearly always better
Does not work
- Adjective-stacking: "professional, engaging, compelling"
- "Act as a world-class expert in…" — it does nothing measurable
- Asking "are you sure?" instead of checking a source
- One endless conversation covering six unrelated tasks
- Politeness padding — courtesy is fine, it just is not doing anything
Turning a Good Result Into a Process
This is the step almost everybody skips, and it is where the actual value is. A prompt that worked brilliantly once, typed into a chat window and then lost, has produced one document. The same prompt written down as a template with the variable parts marked produces one every week, for anyone on your team, without you.
What "written down" means in practice:
- The exact wording that worked, with placeholders where the material changes
- The example of good output attached, so the standard travels with the template
- A note on what it is for and when not to use it
- Somewhere shared — a doc, a wiki page, a saved project. Not one person's chat history.
- A named person who checks the output and owns the standard
The Short Version
Key takeaways
- Brief it like a capable new hire who knows nothing about your business — because that is exactly what it is.
- Five parts: the task as a verb, who it is for, the material, an example of good, and the constraints.
- Attach an example. It beats every adjective and every clever phrasing.
- Tell it to flag gaps rather than fill them. That single line prevents invented figures.
- Write the working prompt down as a template, or you have produced one document instead of a process.
What makes a good AI prompt?
Five things: the task stated as a verb, who the output is for, the material the model needs, an example of what good looks like, and the constraints including what not to do. The example is the highest-leverage of the five — it conveys a standard that adjectives cannot.
Do I need to learn prompt engineering?
Not as a discipline. The useful version is a short checklist you apply while briefing, and it is much closer to management than to engineering. If you can write a clear brief for a competent freelancer, you already have the skill; you just have to remember that the model has no memory and no context you have not supplied.
Why does AI give me generic, bland output?
Because a thin instruction gives it nothing specific to work with, so it falls back on the average of everything it has read — which is, by construction, generic. The fix is context and an example: paste the real material, attach your best previous version, and say who will read it.
Does saying "act as an expert" improve the answer?
Not measurably, in our experience or in most published testing. Role-play openers were more useful with earlier models. Current ones respond far more to concrete context — the actual document, the actual audience, the actual example — than to being told what to pretend to be.
Should I use one long conversation or start fresh?
Start fresh for each new task. A long conversation drags every earlier message into each new answer, including abandoned attempts, and the whole history is re-sent on every turn so it costs progressively more. Fresh conversations are both cheaper and sharper.
How do I stop AI making up facts and figures?
Give it the real numbers in the prompt, and add an explicit instruction: "Do not invent figures. If a number is missing, mark it as a gap rather than filling it in." Models fill gaps by default because a fluent continuation is what they produce. Naming the alternative behaviour prevents most of it.