If you want to know how to write ChatGPT prompts that produce usable output, follow a four-part skeleton: role, context, task, format. Tell the model who it is, what it needs to know, what to do, and exactly how to present the answer. Add a constraint to cap length and scope, and you have a prompt you can reuse and compare.
This is not theory for us. Honeyb, our AI visibility platform, runs structured buyer prompts through AI engines every day. On 13 July 2026 we also ran a controlled test: 20 buyer prompts, three identical runs each, across ChatGPT, Gemini, Claude and Perplexity via API, 240 answers in total. Structure is what makes those answers comparable and machine-parseable. The same principles apply to any prompt you write by hand.
How to Write ChatGPT Prompts: The Four-Part Skeleton
Role sets the model's perspective before it starts generating. "You are a procurement analyst" produces a different vocabulary, depth and set of assumptions than no role at all. Context supplies the facts the model cannot guess: your company size, budget, audience, or the document it should work from. Most weak prompts fail here, because the model fills missing context with the most generic assumption available.
Task states the single action you want, phrased as an instruction rather than a topic. Format tells the model what the output should look like: a numbered list, a table, a 100-word summary, a JSON object. The fifth element, constraint, is the one most people skip. It bounds length, item count and scope, which is what stops answers drifting into essays.
| Component | Weak prompt | Structured prompt |
|---|---|---|
| Role | None, jumps straight to the question | "You are a procurement analyst advising a services firm." |
| Context | "Tell me about project management tools." | "We have 40 staff, bill hourly and need built-in time tracking." |
| Task | Implied, so the model guesses | "Recommend the strongest tools for this firm." |
| Format | Unspecified, so you get a long essay | "Numbered list, most to least recommended, one sentence of reasoning per tool." |
| Constraint | None, so length and scope drift | "Maximum five tools. No preamble, no closing summary." |
The biggest single improvement usually comes from the last two rows. Role and context shape what the model says. Format and constraint shape whether you can actually use it, skim it, or feed it into another system.
A Worked Example: Our Structured-Output Pattern
Honeyb scans buyer prompts daily and compares answers across runs, engines and days. That only works if every answer arrives in the same shape. Here is the pattern we use, adapted for a manual buyer-research prompt.
``` You are a software buyer researching project management tools.
Context: mid-sized services firm, 40 staff, bills hourly, needs built-in time tracking.
Task: recommend the tools you consider the strongest options for this firm.
Want to see this in action?
See how every major AI model talks about your brand. Free to start.
Format: a numbered list, ordered from most to least recommended. One sentence per tool explaining why.
Constraints: maximum five tools. No preamble and no summary after the list. ```
Three moves do the work. First, the format is explicit, so the answer arrives as a list rather than prose. Second, length is constrained, so the model commits to a shortlist instead of hedging with fifteen options. Third, the ordering instruction forces a preference, which is the part you actually want. In our 13 July 2026 test, engines returned between 4.8 and 5.2 brands per answer under this pattern, tight enough to compare run against run.
Structure Controls Format, Not the Answer
Here is the honest limit. A structured prompt guarantees the shape of the output, not its content. In our test, identical prompts run back to back changed their top recommendation 44% of the time on Gemini, 43% on Perplexity, 35% on ChatGPT and 28% on Claude. The list format held every time. The name at position one did not.
Top-pick change rate
How often the top recommendation changes between identical runs
Broader research points the same way: SparkToro found the same AI query changes its answer roughly 70% of the time. Even the full set of brands mentioned only overlapped 42% between identical ChatGPT runs in our data, rising to 67% on Claude. So if your carefully structured prompt returns a different winner tomorrow, the prompt is not broken. That is sampling behaviour, and we cover it in detail in does ChatGPT give the same answer to everyone.
A Checklist Before You Press Enter
- Role: one sentence naming the perspective. - Context: every fact the model cannot infer, stated plainly. - Task: a single instruction, not a topic. - Format: the exact output shape, named. - Constraint: a hard cap on length or item count. - Repetition: if the answer matters, run it more than once before acting on it.
The skeleton pays off twice. It makes your own drafting and research prompts sharper, and if you write longer material with AI, the same discipline applies, as we cover in ChatGPT for content creation. It also matters in reverse: buyers are writing prompts like the worked example above about your category right now, and the 28% to 44% run-to-run variance we measured means no single answer tells you where you stand. See how to measure AI share of voice for the measurement side, or start with a free AI visibility check to see how AI engines answer buyer prompts about your brand today.
