Structuring Complex Prompts

You’ve now mastered the building blocks of prompting. But as soon as the task grows—a ten-page document, a multi-criteria analysis, an entire project—a prompt written in a single stream shows its limits. You need to move to architecture: a structured prompt guides ChatGPT like a statement of work guides a vendor, section by section, with explicit expectations for each.

The Problem with Monolithic Prompts

Here is the typical prompt from a user who knows the basics but doesn’t yet structure their requests. The role is there, the numbers are there, and the requests are there too.

You are a senior marketing consultant. Analyze the digital strategy of our
B2B SaaS company. We have a website with 15,000 visitors/month, a
blog that publishes 2 articles per week, a LinkedIn presence, and a few
Google Ads campaigns. Our CAC is €350 and our LTV is €2,800.
Give me a complete diagnosis with prioritized recommendations and a
6-month action plan with an estimated budget and KPIs to track.

All the ingredients are present, but mixed into a single block of text. The response will be long, disorganized, and will lose half the requests along the way: the diagnosis will take up three-quarters of the output, the budget will be glossed over, and the KPIs will disappear. This isn’t a weakness of the model; it’s a direct consequence of the prompt—nothing distinguishes background information from a deliverable to produce, nor a secondary deliverable from a core one.

Section-Based Structure

Take exactly the same material and distribute it into named sections. The context becomes referenceable, each deliverable is isolated and counted, and the output format is decided upfront rather than left to chance.

## Role
You are a B2B digital strategy consultant with 10 years of experience
working with SaaS startups.

## Context
- Company: B2B SaaS, project management, French market
- Website: 15,000 visitors/month (60% organic, 25% paid, 15% direct)
- Blog: 2 articles/week, traffic growing 8%/month
- Social: LinkedIn only, 3,500 followers, average engagement 2.1%
- Ads: Google Ads, budget €4,000/month
- Current CAC: €350
- LTV: €2,800
- Goal: Reduce CAC to €200 in 6 months

## Requested Analysis
1. Diagnosis of each channel (SEO, content, social, paid) with a score out of 10
2. Identification of the 3 biggest improvement opportunities
3. Identification of the 2 actions to stop immediately

## Action Plan
- 6 months, broken into 3 phases of 2 months
- For each phase: objectives, actions, budget, KPIs
- Prioritization by impact/effort ratio

## Output Format
- Diagnosis: table by channel
- Opportunities and actions to stop: bulleted list with reasoning
- Action plan: one table per phase

The resulting response is generally more complete and better organized, with every request finding its place. One caveat, though: architecture doesn’t replace content. An empty section produces nothing, and a fifteen-section prompt is almost always a prompt that needs consolidation—you likely have two or three distinct requests disguised as one.

Separators That Work

OpenAI’s guide recommends using delimiters, such as Markdown headers, XML tags, or section titles, to clearly indicate the different parts of a prompt. Three families work well. Markdown is the most common, and it remains readable in the conversation history.

## Section 1
content

## Section 2
content

XML tags delimit blocks more strictly. They become valuable when you paste a long document inside the prompt: the closing tag indicates unambiguously where the material ends and where your instructions resume.

<context>
background information
</context>

<instructions>
what ChatGPT must do
</instructions>

<constraints>
rules to follow
</constraints>

Textual separator lines provide the same service with a cleaner look on screen.

--- CONTEXT ---
information

--- INSTRUCTIONS ---
task to perform

--- FORMAT ---
expected format

Choose the one that feels natural to you and stick with it from prompt to prompt: consistency in your own habits beats the theoretical superiority of any syntax.

The “Professional Brief” Structure

For complex tasks, adopt the skeleton of the brief you would hand to an agency. It covers everything a vendor would ask before accepting the work, including the item rarely formulated: how will we know it succeeded.

## Objective
What do we want to achieve? (one sentence)

## Target Audience
Who will read/use the result?

## Context
Essential background information

## Expected Deliverables
Precise list of what ChatGPT must produce

## Constraints
What to respect and what to avoid

## Examples
One or two examples of the expected result (if relevant)

## Success Criteria
How do we know the result is good?

Filled out for a real case, this template yields a long but perfectly readable prompt. Note the “Success Criteria” section: it serves as a guardrail for the model, which will check its timing, as much as for you, so you know what to verify during review.

## Objective
Create a 45-minute webinar script on the theme "Automating
B2B Prospecting with ChatGPT".

## Target Audience
Sales directors and SDRs at tech SMEs, ages 30–45,
familiar with digital tools but beginners in AI.

## Context
Free webinar to generate qualified leads.
150 registrants expected. The business goal is to convert
10% to our paid course.

## Expected Deliverables
1. Catchy title and subtitle
2. Detailed outline with timing per section
3. Word-for-word script for the introduction (3 minutes)
4. 5 key slides with content for each slide
5. 3 moments of audience interaction (polls, questions)
6. Script for the final pitch to the course (2 minutes)
7. Post-webinar follow-up email

## Constraints
- Dynamic and concrete tone, not academic
- At least 3 live demonstrations of ChatGPT prompts
- No AI jargon (no "LLM", "tokens", "fine-tuning")
- The sales pitch must not exceed 2 minutes
- Include a limited-time offer for webinar attendees

## Success Criteria
- The outline fits in 45 minutes (check timing)
- Each section provides concrete, actionable value
- The transition to the pitch feels natural, not forced

Here is the blank version to keep. Pull out the heaviest request you’ve given ChatGPT this month and rewrite it in this template: for a task of this size, twenty structured lines are generally better than five messy ones, and the time spent filling out the sections pays off in reduced corrections.

## Objective
[In one sentence, what do you want to achieve?]

## Target Audience
[Who will use the result?]

## Context
[Essential information, as bullet points]

## Expected Deliverables
[List precisely each element to produce]

## Constraints
[What to do and what not to do]

## Output Format
[How the response should be organized]

The Numbered Subtask Technique

When a single request contains several independent deliverables, number them and require clear separation in the response. This is the simplest fix for forgetting an instruction buried in the middle of the prompt, and it lets you later pick up a single task to deepen without restarting the whole thing: “expand on task 3” is enough; the model knows what you mean.

Complete the following 4 tasks in order.
Clearly separate each task in your response.

Task 1: Summarize this article in 5 key bullet points.
Task 2: Identify the 3 weakest arguments in the article.
Task 3: Write a counter-argument in 200 words.
Task 4: Propose 5 questions to ask the author in an interview.

Article:
[paste the article here]

Long Tasks with ChatGPT Work

Large requests often find their place in ChatGPT Work, the experience designed for tasks that draw on multiple sources or tools, chain steps together, or produce a significant deliverable. The ChatGPT documentation offers advice that extends the brief: describe the result, provide the sources, name the audience, and say how you will review the work. It also recommends starting with a result you can review, limiting sources and the time period to what is useful, separating mandatory work from optional improvements, asking for an outline when the method matters, and requiring your approval before any sending, publishing, or modification of information that others depend on.

Your Turn

Restructure your heaviest request. Take the longest request you’ve made to ChatGPT this month. Rewrite it using the brief template, with separate sections, success criteria, and numbered subtasks. If the task fits and your plan allows it, run it in ChatGPT Work by asking for an outline first, and compare it with the original version.

Official sources: Prompting (ChatGPT documentation) · Reasoning best practices · Prompt engineering (OpenAI API documentation)

Key Takeaways

• As soon as the task grows, split the request into named sections
• Markdown, XML, or textual delimiters clearly separate parts of the prompt; keep the same syntax from prompt to prompt
• The professional brief covers objective, audience, context, deliverables, constraints, and success criteria
• Number subtasks so none get lost, then tackle them one by one
• For a long task in ChatGPT Work, ask for an outline and keep control over sending and publishing