How to use ChatGPT at work: write a work instruction, not a prompt

What self-taught ChatGPT use looks like

Our ChatGPT training page says it plainly: most self-taught users write short prompts and accept the first answer. That works for a one-off question. It does not work for the weekly report, the customer reply or the meeting summary, because next week the same person writes a slightly different short prompt and gets a slightly different answer, and nobody else on the team can repeat it.

Learning how to use ChatGPT at work is mostly learning to write down what you already know about the task, once, in a form ChatGPT can follow every time. We call that a work instruction. It is the first module of our ChatGPT sessions, and every person practises it on a task from their own week.

The five parts of a work instruction

The same five parts, every time:

1. Task. What the output is and who it is for. “Draft the weekly operations update for the leadership team” beats “summarize this”.
2. Context. What the reader already knows, what they care about, and what has gone wrong before. One or two sentences.
3. Source. The material ChatGPT must work from, and an instruction to use only that material. An export, a policy document, last week’s approved version.
4. Format. Headings, length, order, tone. Paste an approved past example if you have one; it does more than a paragraph of adjectives.
5. Check. What ChatGPT should flag instead of guessing, and what the person will verify before the output is used.

Here is what that looks like for one of the workflows on our ChatGPT page, the operations team’s weekly report from a spreadsheet export:

Task: Write this week’s operations update for the leadership team, under 250 words.
Context: Leadership reads it on Monday morning and wants changes against last week, not totals they already know.
Source: Use only the attached export and last week’s update. Do not add numbers that are not in the export.
Format: Three headings: What changed, What is at risk, What we need decided. Same order every week.
Check: List every number you used with the cell or row it came from. If a figure looks wrong or is missing, say so instead of filling it in.

The last line matters most. It turns “is this right?” into a list the person can tick off against the spreadsheet in a few minutes.

How to turn one task into a team workflow

1. Pick one task you repeat every week. Not the hardest part of the job, the most repetitive. If you are unsure which, the start-with-one-task method scores candidates on four questions.
2. Write the five parts in a plain document first. If you cannot fill in the source and check lines, the task is not ready for ChatGPT yet.
3. Run it three times on real work. Each time, note what you had to fix by hand. Most fixes belong in the context or format line, not in a longer prompt.
4. Put it in a shared Project. OpenAI’s Projects in ChatGPT article describes a project as a place for files, instructions and related chats, and shared projects let colleagues work from the same files and instructions. A new hire then starts from the approved version, not from a blank chat.
5. Write down who checks. One name per workflow. The AI workflow one-pager has a line for it, next to the input, the output and the review step.
6. Review after two weeks. Was it used on real work, was the output good enough, and what still needed fixing? The two-week workflow review is that conversation on one page.

Which data goes in

Agree three lists before anyone pastes company material: never goes in (passwords, ID numbers, health data, anything a contract forbids), check first, and fine to use. Our AI data boundaries one-pager is a template for those lists.

Then use a company workspace for company work. OpenAI’s business data page states that it does not train its models on data from ChatGPT Business, Enterprise and its other business products by default. A personal account is governed by that person’s own settings, which is one reason the AI tool approval checklist starts by choosing one approved plan. Terms change, so read your plan’s current pages on the day you decide.

One more thing to know if your team built custom GPTs: OpenAI’s ChatGPT release notes announced plans to retire custom GPTs and offer a migration path. The dates reported differ by workspace, so check the notice in your own workspace. A work instruction written as plain text in a document moves to whatever replaces them without rework.

When this does not work

It does not work on tasks that never repeat. A one-off strategy question does not need a work instruction, and writing one is wasted effort.

It does not work when the source is the problem. If the export is wrong or last week’s approved version does not exist, ChatGPT will produce a tidy summary of bad material. Fix the input first.

And it does not work without the check. A work instruction makes the output more consistent; it does not make it correct. If nobody on the team can tell a right answer from a wrong one for this task, pick a different task. Where the job is really connecting several systems, our ChatGPT vs AI agents guide covers when an agent is worth the extra risk.

What to do in the next 14 days

• Pick one weekly task and write its five-part instruction in a document. Twenty minutes.
• Run it on real work three times and note every manual fix.
• Move the instruction and its approved files into a shared Project on a company account, and name the person who checks the output.

The approach in this post is what we teach in ChatGPT training for employees.

• ChatGPT training for employees
• Start AI training with one task
• AI workflow one-pager
• ChatGPT vs AI agents

Questions people ask

Sources

• OpenAI Help Center: Projects in ChatGPT
• OpenAI: Business data privacy, security, and compliance
• OpenAI Help Center: ChatGPT release notes