GPT-5.6 and ChatGPT Work: how to prompt it, and the playbook

What ChatGPT Work is

The name says the target: the tasks on your job description. However the feature list reads the week you visit, every setting it exposes is a place to put context once, so you stop pasting it into each chat.

Operating software, clicking through a CRM or a form, is a different model's job. My GPT-6 Astra guide covers briefing the computer-use model; this page covers prompting the one that writes, plans, summarizes and analyzes.

How to prompt GPT-5.6

Five habits carry across every GPT model I have prompted. What 5.6 changes in the details is in the guide.

1. Lead with the finished state. "A 400-word onboarding email a new customer reads in one sitting" beats "write an onboarding email".
2. Give it the context it cannot guess. Who reads the output, what has already been tried, what the result plugs into, and what tone the reader expects.
3. Fix the format before it starts. Headings, word count, table columns, file type. A model with no format picks one, and it will rarely pick yours.
4. Ask it to ask. "Before you start, list the questions you need answered, then wait." One line, and the first draft stops being a guess.
5. Show one good example and one bad one. Two examples beat a paragraph of adjectives.

Put together, a prompt for 5.6 looks like this:

Goal: a 400-word onboarding email for a customer who signed up for the annual plan today.
Reader: a marketing manager who has used a competitor and is skeptical.
Context: the four things they must do in week one are in the attached doc.
Format: subject line, four short paragraphs, one link, plain text.
Rules: no exclamation marks, no discount offers, do not mention the competitor.
Before you write, ask me anything you are missing.

The ChatGPT section of the prompt library has this shape applied to job after job.

The agentic playbook

A prompt gets you one answer. A play gets you a job done, and a playbook is the set of plays you run every week without rewriting them.

Each play has five parts: a trigger (Monday at nine, a new row in the sheet, a client email, a calendar event ending), the steps in order, the tools each step may touch, a checkpoint where the model shows its work before continuing, and a done rule that says what finished looks like. Write the play once, run it in ChatGPT Work with the right connectors, and read the checkpoint.

An example play, weekly pipeline review: pull the deals updated in the last seven days, flag any with no next step, draft a two-line nudge per flagged deal to its owner, show me the list before anything is sent. Four steps, one checkpoint, one done rule. That structure is the whole idea. The playbook in the download has written out, plus a prompt library sized for ChatGPT Work.

A play you run every week deserves a permanent home. A custom GPT holds the role, the context to ask for, the frameworks and the output format, so the play starts the same way each time; the mega-instructions in the Custom GPT Toolkit are built exactly for that.

What is inside the GPT-5.6 guide

• GPT-5.6 Prompt Engineering Guide
• ChatGPT Work Agentic Playbook plus Prompt Library
• GPT-5.6 Prompt Engineering Mini-Course
• Lifetime access and updates

It is free. The form sends it to your inbox, and the same address receives the updates as the model changes.