How to Write Better ChatGPT Prompts for Work: 4 Elements

A vague ChatGPT prompt gives you a vague draft you have to rewrite anyway. Learning how to write better ChatGPT prompts for work is less about clever tricks and more about filling in four slots the model needs before it can help you.

In our live workshop we rewrite prompts on real work participants bring in: emails, contracts, spreadsheets, briefs. The same four fixes come up. This post is the short version: the four elements, three before-and-after rewrites, and a checklist you can use today.

The four elements of a working prompt

Every named framework (RTF, CRISPE, RASCEF) rearranges the same four slots. OpenAI's consumer-facing guidance names four parts: goal, context, output, and boundaries. Anthropic's prompting docs, Zapier's prompt guide, Harvard University IT's AI resources for staff, and MIT Sloan's teaching materials teach the same skeleton to non-technical audiences.

Role. Who the model should sound like. "You are a senior contracts attorney reviewing on behalf of the vendor" produces different output than "You are a paralegal drafting for internal use." Pick one.

Context. What the model does not yet know. Who this is for, what has already happened, what constraints matter. Put it at the top of the prompt. As of September 2026, GPT-5, Claude Sonnet 4.6, and Gemini 2.5 all handle long context reliably; the failure mode is missing context, not too much of it.

Task. One clear deliverable. Not "help me with the intro email." A specific output: "Draft the first cold email in a three-touch sequence to the operations lead at a 200-person architecture firm."

Format and constraints. Length, structure (bullets, table, JSON), tone, and what to avoid. This is where you kill the AI voice before it starts.

If a prompt is falling flat, one of these four is usually thin.

Rewrite 1: a follow-up email

Before:

Write a follow-up email to a client who hasn't replied.

Result: a generic "Just checking in!" note that any reader will recognize as AI-drafted.

After:

Role: You are me, a management consultant writing to a warm-but-slow client.
Context: I sent a proposal 9 days ago for a $28k discovery engagement.
The client (COO, mid-sized manufacturing) said in our last call she wanted
to move fast, but has been silent since. She has budget authority. Our
tone with her has been direct and dry, not chatty.
Task: Draft one follow-up email, second touch, that gives her an easy
out and a low-friction next step (30-min working call).
Format:
- Subject line under 45 characters, no question marks.
- 90 to 120 words, plain text, no bullet lists.
- End with two proposed times, one morning and one afternoon.
Avoid: "just checking in," "circling back," "hope you're well."

What changed: the model now has role, stakes, tone, format, and a short list of phrases that would give it away.

Rewrite 2: a contract read-through

Before:

Review this contract and tell me if it's ok.

Result: a summary and a hedge. No decisions you can act on.

After:

Role: You are a commercial-contracts reviewer flagging risks for the
vendor (us), not for the client.
Context: This is a SaaS master services agreement. Deal size $60k/year,
2-year term. Our standard positions: mutual limitation of liability
capped at 12 months of fees; no auto-renewal without written notice;
IP we create stays ours unless separately licensed.
Task: Read the pasted MSA and produce a redline checklist.
Format: A table with four columns:
  Clause number | Quoted language (<=25 words) | Why it's a risk for us
  | Proposed change.
Only include clauses that deviate from our standard positions. If a
term looks acceptable, do not list it.
Constraints: Do not draft new legal language beyond the "Proposed
change" cell. If you are uncertain about a clause, say so in-line.

One note before you paste: do not paste real contracts with counterparty names, pricing, or personal data into a personal ChatGPT account. Redact first, or use a business-tier plan with a data processing agreement. More on that below.

Rewrite 3: a messy spreadsheet

Before:

Look at this data and tell me what's interesting.

Result: platitudes about the data being "diverse" and "rich."

After:

Role: You are a revenue analyst who reports to a sales director.
Context: The attached CSV is 8 months of closed-won deals with columns:
date, ARR, segment (SMB/Mid/Ent), source, sales rep, days-to-close.
The director suspects Mid-market cycles are getting longer and wants a
one-page read before Monday's pipeline review.
Task: Answer three questions with the data:
  1. Is Mid-market days-to-close trending up? Compare the last 90 days
     against the prior 90 days.
  2. Which source has the highest ARR-per-deal in the last quarter?
  3. Are any reps underperforming their own 6-month average by more
     than 20%?
Format:
- One short paragraph per question, 60 words max.
- Numbers rounded to whole dollars or whole days.
- End with a "What I would verify manually" list of 3 items.
Constraints: If the CSV is missing a column you need, say so and stop.
Do not invent data. Show the code you ran.

Two things worth knowing. First, as of September 2026, ChatGPT, Claude, and Gemini can all execute analysis code on uploaded files, and asking for the code back is the only reliable way to check the math. Second, "do not invent data" will not stop hallucinations completely. Spot-check numbers against the source.

The 60-second prompt checklist

Run the prompt through these six questions before you hit enter. It takes about a minute and catches most of the sludge. If the answer still comes back flat, the usual cause is in Why ChatGPT Gives Generic Answers and How to Fix It. To stop retyping the role and context in every new chat, see ChatGPT Custom Instructions Examples for Work: 7 You Can Copy.

#	Ask yourself	If no...
1	Have I given the model a role with a clear point of view?	Add one line: "You are ___ writing for ___."
2	Would a smart new hire have enough context to do this without asking me follow-up questions?	Add the missing background.
3	Is the task one specific deliverable I can copy-paste when it is done?	Split it. One prompt per output.
4	Have I specified format: length, structure, tone?	Add "Return a 5-row table with columns X, Y, Z."
5	Have I named what to avoid: filler phrases, hedging, specific words?	Add an "Avoid:" line.
6	Would I still be comfortable if my client, employer, or a journalist saw what I pasted in?	Redact, or move to a business-tier account.

Where this breaks

Prompt quality alone does not fix a few real problems.

Confidential data. OpenAI's business tiers (ChatGPT Business, Enterprise, and the API) do not use your content to train models by default as of September 2026; personal Free and Plus plans do, unless you turn training off in settings. Even with training off, do not paste signed contracts, client PII, medical or financial records, or anything under an NDA into a personal account. Redact first, or use a business plan with a data processing agreement in place. Anthropic and Google offer similar business tiers for Claude and Gemini. The full list of what should never go into a personal account is in Is It Safe to Upload Confidential Documents to ChatGPT?.

Numbers you cannot verify. All major models still make up figures when they do not know. For anything quantitative, ask for the working (the code, the source, the calculation) and check it. The checks we run before sending anything with figures are in How to Fact-Check AI Answers: A Hallucination Checklist.

One-shot magic. A good prompt gets you most of the way there. The remaining fix is almost always a short follow-up: "Rewrite paragraph two in a shorter, drier register" beats rewriting from scratch. Eight follow-ups that fix most almost-right drafts are in Follow-Up Prompts to Improve ChatGPT Answers: Diagnose First.

Where we go deeper

The first session of our two-week live workshop opens with exactly this: the four elements, then live rewrites of prompts participants bring from their own week. An email that came back flat, a contract clause the model summarized without thinking, a spreadsheet it refused to touch. You leave that session with your own prompt library started, not a folder of generic templates. Sessions two through four extend the same pattern to writing and correspondence, to documents and data, and to your own assistants and no-code automations.