Writing Prompts for Real Work, Not Demos
Share
The prompt that produces an impressive demo and the prompt that survives a real working week are rarely the same prompt.
Direct answer
A work prompt needs four things a demo prompt does not. Your organisation's constraints, the actual source material, a defined output format someone downstream can use, and a check step. Add those and output stops needing a rewrite.
Based on the AI and Automation Practitioner standard (ST1512), a Level 4 apprenticeship for people already in work. Funding rules described here are the England position and are set by the funding body.
Potentially free for you
The learner never pays. Employers potentially pay nothing either, and where the apprentice is aged 16 to 24 the government funds the whole cost even if the employer does not pay the levy.
Closing any gap. A levy transfer from another employer covers 100% of the cost. Unused transfer allowance is not carried forward at the end of the tax year, and it must be agreed before the apprentice starts.
From 1 August 2026 the government stopped adding the 10% top-up to new funds entering levy accounts, so there is less in the system than there was.
Part of our guide to Prompt Engineering: What It Is and How to Learn It.
Why demo prompts fail at work
A demo optimises for surprise. Real work optimises for consistency, because the output goes into a process that expects a shape.
Demos use generic inputs. Real work uses your customers, your policies and your data, none of which the model has seen.
Demos are judged once. Work prompts run forty times a week, and the failure rate that looks acceptable on one output is not acceptable across forty.
That is why people report a great first experience and quiet abandonment two weeks later.
Turning a vague task into a brief
Write down what a good output looks like before you open the tool. If you cannot describe it, the tool cannot produce it.
Name the reader. Internal colleague, customer, regulator or supplier. Each one changes the register entirely.
List what must never appear. Prices you cannot confirm, commitments you cannot make, tone you do not use.
Decide the format. Bullet list, three paragraphs, a table with named columns. Downstream processes care about shape.
Building in the check
Ask the output to flag anything it was unsure about. It will not always comply, and when it does it saves you time.
For numerical work, ask for the calculation alongside the answer. A wrong number with visible working is caught in seconds.
Ask for the source of any factual claim. Claims with no source attached are the ones to verify first.
Build a two-line checklist for whoever reviews the output. Unowned checks do not happen.
There is more on this in our guide to Prompt engineering explained.
Making it survive a real week
Run it on ten real cases from last month where you already know the right answer. That is the only honest test.
Count the ones that needed heavy rewriting. If it is more than two in ten, the prompt is not ready.
Save the version that worked with a one-line note about what it is for and when you last checked it.
Review it quarterly. Models change and prompts carrying heavy scaffolding often work better simplified later.
Writing for the person downstream
Most work output is not the end of a process. It goes to a colleague, a system or a customer with expectations of its own.
Ask what the next person needs before you decide the format. A summary that reads well but lacks the reference number creates work rather than saving it.
Where output feeds a spreadsheet or a case system, specify the exact column names and order. Generic prose then has to be retyped.
This is the difference between a prompt that saves you time and one that moves your work onto somebody else.
The tasks worth doing this for
High frequency and moderate stakes. Something you do weekly where a mistake is embarrassing rather than expensive.
Text-heavy with a predictable shape. First-line replies, meeting summaries, document drafts, data reformatting.
Work where a competent person can check the output quickly. If checking takes as long as doing, the saving disappears.
Avoid anything safety-critical or legally binding until you have run it against real historic cases and measured the failure rate.
Key takeaways
- Describe a good output before you open the tool.
- Name the reader, because it changes everything about the register.
- Test on ten real cases where you already know the answer.
- More than two heavy rewrites in ten means the prompt is not ready.
- Pick tasks where checking is quicker than doing.
Frequently asked questions
Why does my prompt work once then fail?
Demo prompts are judged on one output. Work prompts run repeatedly, and a failure rate that looks fine once becomes obvious across forty runs.
How do I test a prompt properly?
Run it on ten real cases from last month where you already know the right answer, and count how many need heavy rewriting.
What makes output usable rather than impressive?
A defined format, your organisation's constraints, the actual source material, and a check step someone owns.
Should I ask AI to flag its own uncertainty?
Yes. It does not always comply, but when it does it tells you which parts to verify first, which saves real time.
Which tasks are worth prompting?
High frequency, moderate stakes, text-heavy work with a predictable shape, where a competent person can check the output quickly.
Which tasks should I avoid?
Anything safety-critical or legally binding, until you have measured the failure rate against real historic cases.
How often should I review a saved prompt?
Quarterly. Models change, and prompts carrying heavy scaffolding often work better simplified after an upgrade.
Who should check the output?
A named person, not a principle. Unowned checks do not happen, which is how errors reach customers.
The funded route, in figures
Figures are the England position. Levy-paying employers usually draw the whole cost from the levy they already contribute, so they pay nothing extra. Employers below the £3 million pay bill threshold pay nothing for an apprentice aged 16 to 24. For an older apprentice the government funds 95 per cent, and a levy transfer from a larger employer can cover the rest. Eligibility and funding rules are set by the funding body and can change.
Where to go next
If you want this work counted towards a qualification, the funded Level 4 route assesses exactly these improvements in your own job.
We help UK employers and employees access government-funded AI training. Training is delivered by third-party providers. We do not guarantee a place, a qualification, a job or any particular outcome. Eligibility and funding rules are set by the funding body and can change. Always check the current position before you commit. See our disclaimer.