Prompts for work tasks: five parts and six examples for ChatGPT and Claude

One sentence

Ask“write a reply to the customer”Modelguesses the contextGeneric answerrewrite by hand

Six parts

Promptrole, data, example, formatModelsees everything it needsDraftminor edits

The model is the same in both cases. The result depends on what the prompt says: six parts that recur in the Anthropic, OpenAI and Google guides.

Human-in-the-loop: the path of one request

Takeaway

In this exampleA person holds every key point: sets the task, builds the prompt, checks the answer, sends it back to refine and makes the decision. The model only answers what made it into the request.

In generalEvery arrow has its own checkpoint: data goes in whole, the answer is checked against who needs it, and a good request is saved as a template so the next task of the same kind starts from it.

Next stepBefore writing the request, name the reader and the decision they will make from the answer: the role and the format follow from that.

Next step · HDDHarness-Driven DevelopmentHow to turn the check in this diagram into a rule that works for every next task

Five parts of a prompt

Role

Audience and goal

The whole data

Example

<example></example>1–3

Format

5 points

Question last

order

<…>tags

Takeaway

In this exampleFive parts plus tags, each backed by an official guide.

In generalThe model knows only what the prompt says. Anthropic suggests a test: show the prompt to a colleague with no context; if they are confused, the model will be too.

Next stepOpen your last prompt and mark which of the five parts are missing. Add a role and a format: two sentences.

Six examples

Reply to a customer thread

The customer writes for the third time; the tone must stay calm, and only what the thread already says can be promised.

RoleWho and whyFull dataExampleFormat

You are a support agent at [company].
Below is the full thread with the customer.
<thread>
[paste in full]
</thread>
Write the next reply. Tone: calm and friendly, like this example:
<example>
[one good past reply]
</example>
No more than 5 sentences. Mention dates only if they appear in the thread.

Takeaway

In this exampleA role, the thread in tags, a tone sample and two format rules.

In generalA past reply sets the tone better than any adjective. The dates rule is phrased as a permission with a condition, and the model follows it literally.

Next stepKeep two or three of the team’s best replies and drop one into the example block.

Preparing for a meeting

A 30-minute meeting tomorrow; the materials are scattered across threads and docs.

RoleWho and whyFull dataExampleFormat

Tomorrow I meet [whom] about [topic], 30 minutes. My goal is [the decision I need].
<materials>
[threads and documents]
</materials>
Prepare:
1. Three questions to settle at the meeting.
2. For each, what the materials already say, with a short quote.
3. What I need to find out before the meeting.
Format: a three-column table.

OpenAI Academy has an agenda template with time estimates.

Takeaway

In this exampleThe meeting goal, materials in tags and a numbered list of what to return.

In generalThe meeting goal filters the materials. Numbered steps help when order and completeness matter.

Next stepBefore any meeting of 30 minutes or more, run the materials through this prompt and open with its table.

Summary of a long thread

A chat of 200 messages; you need to know what was decided and who does what.

RoleWho and whyFull dataExampleFormat

<thread>
[paste in full]
</thread>
First, copy the exact quotes where decisions are made or tasks are assigned.
Then build a summary from those quotes:
— decisions;
— tasks with owners and dates;
— open questions.
One line per item. If a quote names no owner, write “unassigned”.

Takeaway

In this exampleThe long thread sits on top, the task at the end, quotes first, then the summary.

In generalFor long documents Anthropic advises asking for quotes first and placing the question after the data: in tests that improved quality by up to 30%. The “unassigned” rule stops the model from inventing an owner.

Next stepFor any thread longer than one screen, start by asking for quotes.

Analysing a table

A month of tickets exported to CSV; find what changed.

RoleWho and whyFull dataExampleFormat

Attached is an export of [tickets] for [period], CSV.
Count [tickets] by [topic] and show the 5 biggest changes versus last month.
For each, the number and one line on what may explain it.
Show the code you used so I can check it.

Both ChatGPT and Claude process the file in an isolated sandbox .

Takeaway

In this exampleThe whole file, an exact metric and breakdown, a “top 5” limit and a request for the code.

In generalThe model computes with code in a sandbox, and code can be re-read. OpenAI advises checking the code, output and assumptions before relying on the result.

Next stepCheck one number from the answer by hand. If it matches, trust the rest more.

Text in the company’s voice

You need a post or email that sounds like the team, not like a template.

RoleWho and whyFull dataExampleFormat

Here are three of our posts that landed well:
<example>[post 1]</example>
<example>[post 2]</example>
<example>[post 3]</example>
Write a post about [topic] in the same style: length, tone and structure as in the examples.
For whom: [audience]. Why: [what we want the reader to do].

Takeaway

In this exampleThree examples in tags and one line on who it is for and why.

In generalAnthropic recommends 3–5 examples; Google’s guide also lists examples as the first practice. A style shown in examples is copied more precisely than a style described in words.

Next stepPut good texts into a project and use them as examples, so you do not paste them every time.

Researching a topic

You need an overview of a market or practice with sources, half a day of searching by hand.

RoleWho and whyFull dataExampleFormat

Research: [question].
For whom: [audience]. Why: [the decision to make].
Data no older than [year].
Output: a table [columns] and a 5-point conclusion.
A source link for every fact. If there is no data, say so.

Both chats have a deep research mode: deep research in ChatGPT and Research in Claude .

Takeaway

In this exampleA goal, a freshness limit and an output format with links.

In generalIn research mode the model runs a series of searches and returns a report with links. Links let you check every fact, and “if there is no data” stops gaps being filled with guesses.

Next stepOpen two or three links from the report and check the fact against the source before passing the conclusion on.