By now you know the recipe (role, task, context, examples, format) and you know how to steer a reply. This lesson is about noticing that different kinds of work want slightly different prompts. The recipe never changes underneath. But a prompt for an email leans on tone and examples, while a prompt for a spreadsheet leans on the exact question you're asking.
Here are four patterns for the four jobs you'll hit most: writing, analysis, research, and coding. Each comes with a worked example. Steal them, adapt them, keep the ones that fit your week.
Writing: drafts, edits, and tone
Writing tasks live and die on two things: who's reading, and what it should sound like. Lead with those. The biggest upgrade is a sample of the voice you want instead of a description. "Match this" beats five adjectives.
You're writing as me: [who you are, in a line]. The reader is [who], and they care about [what].
Task: [the one thing to write].
Tone: like this sample [paste something you've written]. Not stiff, not salesy.
Format: [length and shape]. End with [the ask or next step].
Worked example, a LinkedIn post about a product update:
You're writing as me, a founder who hates LinkedIn cringe. The reader is a busy operations manager scrolling on their phone.
Task: a short post announcing we cut onboarding from two weeks to two days.
Tone: like this sample [paste a post of your own]. Dry, human, no "thrilled to announce."
Format: under 100 words, one clear point, a line break before the last line. No hashtags.
The draft lands in a writing block. If it's off, steer it with one note (that's the last lesson).
Analysis: turning data and text into findings
The mistake is "analyse this" and hoping ChatGPT guesses what you're after. Analysis prompts need a real question. Attach the file; for spreadsheets ChatGPT can run real calculations (it writes and runs code behind the scenes) and build tables or charts.
Here's [the data or text]. [Attach it.]
Question: [the specific thing you want to know].
Look for: [trends, outliers, the why behind a number].
Give me: [top three findings, a short table, one recommendation]. Show how you calculated the numbers, and flag anything you're unsure about rather than guessing.
Worked example, a month of support tickets:
Here are last month's support tickets [attached].
Question: what are people struggling with, and is it getting worse or better than the month before?
Look for: the top themes by volume, and any theme growing fast even if it's still small.
Give me: the top five issues as a table (issue, rough count, trend), then one line on what to fix first. Say so if the data's too thin to call.
Two lines do the heavy lifting. "Show how you calculated" lets you check the numbers. "Flag what you're unsure about" stops a confident-sounding figure you can't trust. Clean data helps too: headers in the first row, one record per row, no merged tables.
Research: gathering and checking sources
Research is where a wrong answer does the most damage, because it sounds just as sure as a right one. The pattern is built on one habit: ask for sources, and check the ones that matter. Use web search for a quick answer, deep research for a full report.
Research [the question]. Search the web.
I want: [what you're after, and how deep].
For each key claim, give me the source, a link, and the date. Separate what you're confident about from what's shaky. If sources disagree, show me.
Worked example, checking a market before a pitch:
Research the market for meal-kit delivery in Belgium. Search the web.
I want: rough market size, the three biggest players, and one trend that changed in the last year.
For each number, give the source, a link, and the date. Flag anything you can't back up. If figures disagree, show the range; don't average them into one tidy fake number.
Then do the part no prompt does for you: open the links on the claims you're about to repeat. ChatGPT is a fast researcher, not a fact-checker.
Coding: explaining, writing, and fixing
You don't need to be a developer. Plenty of business work touches a formula, a script, or a snippet someone left behind. Coding prompts want the language, the goal in plain words, and a sample input and output.
[Language or tool: an Excel formula / a small Python script.]
Goal: [what it should do, in plain words].
Input looks like: [a sample].
Output should look like: [a sample].
Explain what it does in one line so I can check it.
Worked example:
An Excel formula.
Goal: in column D, flag any order over 1000 euros that's still unpaid.
Input: column B is the amount, column C is the status ("paid" or "unpaid").
Output: the word "CHASE" if it's over 1000 and unpaid, otherwise blank.
Give me the formula and one line on how it works.
Sample in, sample out: you can check the result against your own sheet in seconds. For bigger software work, ChatGPT has Codex, a Part 5 topic.
Stop repeating yourself
ChatGPT gives you two places to park what you'd otherwise paste every time. Your custom instructions hold the things true for all your work, like your role and preferred tone. A project's instructions hold the things true for one stream of work, like a client's style guide or a standard boundary ("use only the attached price list"). And because ChatGPT may also remember earlier chats, glance at what shaped an answer when something feels off. The prompt then only needs what's new about this task.
The thread running through all four
These are one recipe wearing four hats. Every pattern gives ChatGPT a clear goal, the context it needs, and the shape of the answer. Writing leans on tone and examples. Analysis and research lean on the exact question and honesty about uncertainty. Coding leans on sample input and output. Same bones, different emphasis.
Try it
Pick the goal closest to your work right now: writing, analysis, research, or coding. Take the matching pattern and fill it in with something real from your week. Send it.
If the first answer is close but not there, don't start over. Steer it with one specific follow-up. You've now used every tool in the Prompting part on real work.