Provide Context and Constraints

ChatGPT knows nothing about your situation. It doesn’t know your job, your market, your customers, or what you’ve already tried without success. Without this information, it improvises based on averages, and an average never fits anyone perfectly. Context and constraints are the two levers that turn a “correct” response into a “tailored” one.

Why context changes everything

Ask Write an About page for a website and you’ll get generic text filled with “we are passionate about excellence” phrases we’ve all read a hundred times. The model has no material specific to your company: it fills the void. Give it substance, and the same request produces text no one else could publish because it rests on facts belonging only to you.

Write the "About" page for Greenflow, a 12-person French startup
founded in 2023 that develops carbon footprint measurement software for
SMEs in the industrial sector. Our tone is direct and technical, with no greenwashing.
Our clients are CSR directors and quality managers.
Our differentiator: we integrate directly with ERP data
(SAP, Sage) without manual entry.

The five types of context to master

Identity context answers “who are you.” It establishes your position, sector, and room to maneuver, which immediately shifts the response register: I am a training manager at a 500-employee company in the banking sector. I report to the HR Director. A training manager reporting to an HR Director doesn’t have the same decision-making power as a director, and the model will account for this in its recommendations.

Situational context describes what’s happening right now, with its deadlines and limits—We're launching a new product in 3 weeks. The market is highly competitive and our marketing budget is limited to €15,000. Two lines like this immediately eliminate half the unworkable recommendations: those assuming six months of preparation or ten times the budget.

Audience context specifies who will read the output, and this often determines the technical level: The document will be read by investors with no technical expertise who want to understand our competitive advantage. Data context provides the numerical material ChatGPT must rely on instead of inventing order-of-magnitude estimates.

Here are our figures from last quarter:
- Revenue: €450,000
- New clients: 23
- Churn rate: 4.2%
- NPS: 67

Then there’s history context, the most often forgotten: what you’ve already tried and why it didn’t work. I already tried a Google Ads campaign that cost €3,000 for 12 qualified leads. The cost per lead is too high. Without this detail, ChatGPT will recommend exactly what you just abandoned, and you’ll waste a full turn explaining it to the model. One caveat: provide what affects the response, not your autobiography. If an element doesn’t change the outcome, leave it out.

Where to place context

Context doesn’t always come through the prompt text. ChatGPT’s documentation recommends sharing only information that can change the result, and stating what ChatGPT should draw from each source. It distinguishes several channels:

• Attachments: documents, spreadsheets, presentations, or PDFs to summarize, compare, or transform, and images when the task depends on a visual; in that case, show the relevant area instead of letting the image speak alone.
• Web search, when the response depends on recent information; ask for sources so you can verify them.
• Projects, when multiple related conversations need to share the same files and sources.
• Connected sources, like Google Drive or Slack, provided the corresponding plugin is installed: name where to look and what to find, without describing each search.
• Personalization: preferences valid for all your conversations go in Custom Instructions, under Settings > Personalization; what concerns only the current request stays in the prompt.

The art of constraints

Constraints don’t restrict ChatGPT; they guide it. Without them, it applies its default settings, which are rarely yours: too long, too smooth, too enthusiastic. Format constraints frame length and structure (Respond in 200 words maximum, Use bullet points, Structure in 3 parts: problem, solution, benefit), while tone and style constraints set the voice: professional but accessible, no technical jargon, using formal address, no superlatives or exaggerated promises.

The most underused category remains exclusion constraints, which state what absolutely must not be done. They often fix in one line a flaw that several rewrites hadn’t corrected. However, OpenAI’s help center gives one rule to keep in mind: instead of only saying what not to do, also say what to do instead. An exclusion alone leaves the model to guess the alternative; an exclusion followed by the right practice leaves it no choice.

- Do not mention competitors by name
- Do not use bullet points (write in paragraphs)
- Do not start with "Of course!" or "Absolutely!"
- Do not invent any statistics: flag missing data

Make sure your constraints don’t contradict each other: “be concise” and “detail every point” can’t both be satisfied, and then the model arbitrates for you. Also keep moderation: ChatGPT’s documentation advises focusing on one or two limits that prevent a real problem, like an approved figure not to change or a message to prepare without sending, rather than trying to control every step. In its vocabulary, format constraints relate to output, and exclusions to limits.

Bringing both levers together

Here’s context and constraints applied to a common task. Every line in the “Context” block serves a purpose—the funding round justifies hiring, the technical culture dictates the tone—and the last two constraints eliminate clichés that betray a hastily written job post.

Write a LinkedIn post to announce the hiring of a senior developer.

Context:
- Company: DataPulse, SaaS data analytics publisher, 35 employees, Paris
- Looking for a senior Python/FastAPI developer to strengthen the backend team
- Culture: remote-first, systematic code reviews, continuous deployment
- Raised €4M in Series A 6 months ago

Constraints:
- Tone: casual but professional, reflecting tech culture
- Length: 150-200 words
- Include 3 concrete reasons to join us
- End with a call to action
- Do not use clichés like "join the adventure" or "fast-growing startup"
- Do not mention salary

The “complete brief” technique

For complex tasks, organize your context like the brief you’d hand to an agency: the same sections, in the same order, including the section always forgotten in verbal briefs—the prohibitions.

## Objective
Create a 10-slide sales presentation for our CRM solution.

## Target audience
Sales directors at SMEs (50-200 employees), B2B services sector.
They currently use Excel or a basic CRM.

## Key messages
1. Time savings of 5 hours/week per salesperson
2. Native integration with existing tools (email, calendar, telephony)
3. Measurable ROI in 3 months

## Constraints
- Each slide: one catchy title + 3-4 bullet points maximum
- No technical jargon
- Include a customer testimonial slide, with a placeholder for a real testimonial
- Last slide = call to action for a demo

## What not to do
- No direct comparison with Salesforce or HubSpot
- No unverifiable numerical promises

Keep the following template handy and fill it out for the next professional task that comes your way. The exercise takes a few minutes, and its main merit is forcing you to articulate what you kept implicit—especially the “what NOT to do” line, which is almost always the one that would have saved you the most time.

[Clear instruction]

Context:
- Who I am: [your role, company, sector]
- Situation: [context of the request]
- Audience: [who will read/use the output]
- Key data: [figures, facts, relevant information]

Constraints:
- [Expected format]
- [Tone and style]
- [Length]
- [What NOT to do, and what to do instead]

Your turn

Fill out the template. Take the prompt you wrote in Lesson 1. Complete it with the template above: at least three types of context, including history, and two limits formulated with what to do instead. Then ask yourself whether part of this context would be better placed elsewhere: an attachment, a project, or your Custom Instructions.

Official sources: Prompting (ChatGPT documentation) · Best practices for prompt engineering with the OpenAI API · How do I create a good prompt for an AI model? · ChatGPT Custom Instructions (OpenAI help center)

Key takeaways

• Context lets ChatGPT adapt its response to your situation: identity, situation, audience, data, and history
• It also comes through attachments, web search, projects, connected sources, and Custom Instructions
• Constraints guide form and tone; one or two well-chosen limits beat a long list
• An exclusion works better when it also says what to do instead
• For a complex task, structure your request like a professional brief