QWE AI Academy The number-one mistake when writing prompts for ChatGPT? Front-loading the persona and burying the constraints. People open with “You are an expert marketing strategist with 20 years of experience…” and then forget to mention the audience, the word count, or the format. The model has a beautiful costume and no script.
Reverse-engineer it. The thing that actually changes output quality isn’t who ChatGPT pretends to be – it’s how tightly you define what “done” looks like. Constraints first, persona second.
The key takeaway (read this even if you skip the rest)
A better prompt is a prompt with fewer interpretations. Every sentence you write should either add a constraint (“must be under 200 words”) or supply context the model couldn’t guess (“the reader is a non-technical CFO”). Anything else is decoration.
That’s the whole game.
Quick background: why prompts even matter in 2026
ChatGPT runs on large language models that predict the next token based on the words you give it. OpenAI Academy puts it plainly: because LLMs don’t “know” things the way people do, how you phrase your prompt plays a big role in shaping the response – your prompt guides the model so it can generate a helpful reply. Garbage phrasing, garbage output. It’s that direct.
Here’s the wrinkle most guides haven’t caught up to: as of mid-2025, with GPT-5, the model is far better at reasoning through complex instructions and handling longer, multi-step prompts. You can consolidate context and related instructions into one request rather than fragmenting them into many smaller prompts (OpenAI Academy). The old advice (“one task per prompt”) no longer applies to current models.
Method A vs Method B: how to write better prompts for ChatGPT
Two real approaches people use. Both work. One works much better.
| Aspect | Method A: Persona-First | Method B: Constraint-First |
|---|---|---|
| Opens with | “You are a senior copywriter…” | “I need a 150-word LinkedIn post for SaaS founders, formal tone, no emojis.” |
| Strength | Sets vibe and vocabulary | Eliminates ambiguity about the output itself |
| Weakness | Model still guesses length, format, audience | Can feel mechanical until you add a single role line at the end |
| Best for | Brainstorming, exploration | Anything you’ll actually use |
Method B wins almost every time because it removes the model’s incentive to guess. Even with GPT-5’s improved ability to read implicit context, OpenAI’s guidance is consistent: specifying role, audience, or format produces the most accurate and relevant results. Roles still help – they’re just not where you should start.
The constraint-first walkthrough
Here’s the structure. Five lines, in this order, every time:
- Output spec – format, length, structure. “Return a Markdown table with 3 columns: feature, benefit, objection.”
- Audience – who reads this. “For a non-technical hiring manager.”
- Source material – paste the inputs the model needs. Bullet points, raw text, a doc.
- Constraints and exclusions – what NOT to do. “Don’t use the words ‘new’ or ‘solution.’ No bullet lists.”
- Role (optional, at the end) – “Write as a skeptical senior editor.”
A working example:
Write a 120-word product update email.
Audience: existing users who haven't logged in for 30 days.
Features shipped this month:
- New CSV export (any report, any size)
- Dark mode
- Keyboard shortcut for search (Cmd+K)
Constraints: friendly but not chirpy. No exclamation marks. No "we're excited." Subject line first, body second.
Write it as a PM who actually uses the product.Notice how the role line is at the bottom and takes one sentence. That’s deliberate. The model doesn’t need a backstory – it needs a brief.
Pro tip: When a prompt gets longer than ~10 lines, switch to Markdown headers (## Task, ## Context, ## Constraints) or XML-style tags. According to OpenAI’s API documentation, Markdown headers and lists help mark distinct sections and communicate hierarchy to the model, while XML tags delineate where one piece of content begins and ends. This isn’t just for API users – ChatGPT parses these the same way.
Iteration is not optional
Even the best first-shot prompt usually needs one refinement pass. The OpenAI Help Center frames it as an iterative loop: start with an initial prompt, review the response, then refine the prompt based on the output – adjust wording, add context, or simplify as needed.
The trick is to refine surgically. Don’t rewrite the whole prompt – add one constraint, regenerate, compare. “Make it 30% shorter.” “Remove the second paragraph.” “Rewrite this in the same tone but for a technical audience.” One change, one regeneration. You learn what actually moved the needle.
Edge cases nobody tells you about
ChatGPT doesn’t give you a temperature slider
If you’ve read API docs, you’ve seen the advice: for most factual use cases such as data extraction and truthful Q&A, a temperature of 0 is best (per the OpenAI Help Center). The catch: this control only exists in the API. In the ChatGPT app, you can’t set it. Workaround – bake it into the prompt itself: “Be literal. Don’t speculate. If you don’t know, say so.” It’s not a real temperature change, but it gets you closer.
Asking for a word count works better than you’d think
People sometimes assume you have to fight the model on length. You don’t. “Write exactly 150 words” usually lands close to the target. The reason it works in chat is that max_completion_tokens – an API-only parameter – does not control the length of the output. It’s a hard cutoff limit for token generation, and ideally you won’t hit it often because the model stops when it thinks it’s finished (OpenAI Help Center). In chat, the model’s own sense of “done” is the only lever, so telling it the target word count directly is the right move.
The context window is huge – focus matters more than capacity
As of mid-2025, context windows range from around 100k tokens up to one million tokens for GPT-4.1 (OpenAI API docs). That’s a lot of room. But paste a 50-page PDF and ask “what do you think?” and you’ll get something vague enough to be useless. The problem isn’t capacity – it’s that the model doesn’t know what you actually care about. Point it at the right sections: “Focus only on the risk section, pages 12-18. Ignore everything else.” That one instruction changes the answer more than any persona ever will.
Tone words are cheap and underused
One adjective changes a lot. Descriptive adjectives like formal, informal, friendly, professional, humorous, or serious help guide the model, per OpenAI’s guidance. “Skeptical.” “Plainspoken.” “Dry.” Adding one of these is often the difference between a usable draft and a corporate sludge paragraph.
FAQ
Do I need to say “please” and “thank you”?
No. It doesn’t measurably improve outputs. Be polite if you want – the model doesn’t care either way.
Should I still use “Act as a [role]” prompts in 2026?
They still work, but they’re not the magic bullet they were in 2023. With GPT-5, the model already infers a reasonable persona from your task description. Where roles still earn their keep is when you need a perspective the model wouldn’t default to – “act as a skeptical reviewer looking for holes” or “act as a non-native English speaker reading this.” Roles that change the lens, not roles that pad the prompt.
What’s the single biggest improvement I can make today?
Add an explicit output spec to the top of every prompt – format, length, structure. That one habit addresses the most common reason ChatGPT produces vague or off-target responses.
What to do next
Open ChatGPT right now. Take a prompt you used in the last week – anything – and rewrite it constraint-first using the five-line structure above. Run both versions. Compare. That A/B test will teach you more than any guide.