Keywordshow to write better chatgpt promptsBlogai for workai at workworkplace aiai for professionalshow to use ai at workai productivity
Related searcheshow to write better chatgpt promptschatgpt prompts for workprompt engineering for businessprofessional chatgpt promptswork brief for chatgpthow to use how to write better chatgpt prompts
Better ChatGPT prompts start as a brief, not a wish
Most weak ChatGPT output is not a model failure. It is an assignment failure. People write a sentence they would never send a colleague: “make this better,” “write a report,” “analyse this.” ChatGPT still returns a confident page. That confidence is the trap. A better prompt is not clever wording. It is a complete brief: who you are asking it to be, what the artefact is, who will read it, which facts are in play, what it must not do, and how you will know it is done.
Prompt engineering at work is closer to management than to magic phrases. The people who get reliable drafts already know how to assign work. They state the reader. They attach the pack. They say what “done” looks like. They ask for gaps instead of guesses. If you cannot brief a human, you will not brief a model. The difference is that the model will not look confused. It will look finished.
This essay is a practical method for writing better ChatGPT prompts on real workplace tasks: emails, memos, agendas, commentary, and research questions. It is not a list of viral one-liners. One-liners are how you get generic tone and invented numbers.
If you cannot write the reader, the ask, and the source in three lines, do not open the chat yet. You do not know what you are requesting.
The six parts of a prompt that holds up at work
Use six parts every time the output might be forwarded. Role tells the model which craft to imitate: a chief of staff, a financial commentator who refuses to invent figures, an editor who cuts. Task is the artefact and the job it must do. Context is the reader and what they already know. Constraints are the bans: no new numbers, no legal advice, no named-person judgements. Output is the shape and length. Quality checks ask the model to list inferences and missing sources.
You do not need all six for a two-line rewrite of your own paragraph. You do need them when the draft could be mistaken for a source. The more the artefact will travel, the more of the brief you write down. Saving thirty seconds on the prompt is how you spend twenty minutes on a rewrite.
| Part | What to write | Weak version |
|---|---|---|
| Role | The craft and the standard | You are a helpful assistant |
| Task | The artefact and the decision it supports | Write something on Q3 |
| Context | Reader, what they know, what changed | For the team |
| Constraints | Bans, length, data class | Make it professional |
| Output | Shape: memo, table, bullets, gaps marked | A report |
| Quality | List inferences and missing facts | Do a good job |
Iterate on the brief, not on luck
Generate once from a complete brief. Then change one instruction: tone, length, or missing-fact behaviour. If you change five things at once, you will not know what worked. Keep the brief in a shared note so the next person does not start from “write a report.”
Ask for a critique pass as a separate step. “List claims not in the sources. List generic phrases. List what a sceptical reader would challenge.” Then you edit. Do not ask the same model to be the only reviewer of its own charm.
Role: You are an editor preparing a draft I will sign.
Task: Produce the artefact below using only the sources I paste.
Context: Name the reader and what they already know.
Constraints:
- If a fact is not in the sources, write [missing] instead of guessing.
- Separate inferences from stated facts.
- Keep to the word limit I set.
Output: The artefact, then a critique list.
Quality checks: What might be wrong, too generic, or unverified.Save the prompt that survived contact with real work. A library of five briefs your team actually uses is worth more than a hundred screenshots of other people’s prompts.
Prompt habits that quietly fail
The most common failure is the missing source. The prompt says “use the latest numbers” and there are no numbers in the paste. ChatGPT will still write a percentage. The second failure is the missing reader. A board note and a Slack update are different artefacts. The third is the missing ban. If you do not say “do not invent a clause,” you will get a clause.
A subtler failure is over-specifying tone (“punchy, world-class, synergistic”) while under-specifying facts. You will get a brochure. Specify the decision and the evidence. Tone is a last pass, not a substitute for content.
| Mistake | What it looks like | What to do instead |
|---|---|---|
| One-line wish | Make this better | Reader, ask, source, length, bans |
| Hidden research | What is the market size? | Paste the pack or send the person to a source |
| Self-review only | Is this good? | Ask for missing facts, then you check the file |
| Prompt hoarding | A private trick in a chat | A shared brief with an owner |
A prompt that produces a fluent page from no sources is not a good prompt. It is a machine for plausible fiction.
Related reading on StudyGrid
Read next: Professional Prompting How to Use ChatGPT at Work How to Build a Prompt Library. Those essays sit beside this one. Use them when you need the neighbouring skill, not as a substitute for the check you still have to make.
What to do this week
Take one artefact you will send this week. Write a six-part brief before you generate. Generate once. Check claims against the file. Keep the brief if it earned its place. That is how better ChatGPT prompts become a team habit instead of a personal superstition.