OpenAI says keep prompts short. Here's where that advice will hurt you

01 — The context

What exactly does OpenAI advise, and why check it?

The guide's headline message: the model has changed so much that everything you've accumulated and polished in prompts can be thrown away — start from zero.

Sounds convenient. And dangerous — because vendor advice is written for "the average user with the average task," and you're probably not them. I read the guide in full, and there are three theses worth unpacking publicly: where they're right, and where the advice would break what already works for you.

Careful, lots of words))) — but every thesis ends with a practical rule you can take.

02 — Thesis 1

"Prompts should be short now" — true?

OpenAI says: just tell the model WHAT you want, and it will figure out HOW.

First: if the task is simple, universal and one-off — yes, a long prompt is unnecessary. "Translate," "cut to three paragraphs," "explain the term" — any context dance here is wasted.

But if you're, say, a lawyer using an assistant to review contracts — every word matters there. A short prompt won't carry that, and you'll get something average you'll then rewrite by hand for an hour.

Second: if you work on tasks built on your unique expertise rather than average internet templates — how exactly will the model climb into your head and extract it? Spoiler: it won't. Which is why long prompts with context remain everything.

Short prompt — when the internet-average result satisfies you.
Long prompt with context — when the result must be yours:
your expertise, your standards, your phrasing.
Prompt length = cost of error × uniqueness of the task.

03 — Thesis 2

"Never say never": why do bans work poorly?

Here OpenAI is right, and this one's worth taking. When we write bans like "don't write template-style," the model processes the word "template" as a probability branch — and the result gets worse. It's the human "don't think of a pink elephant." You just thought of one, didn't you)

What to do: for every "don't" in your head, ask "then how?" — and write exactly that:

But here's my "but" to OpenAI: what about assistants that need strict no-exception rules? A customer-support assistant that must not leave the script, say. A few bans will still have to be written — just don't build the whole instruction on them. Positive conditions are the foundation; bans are the red lines.

04 — Thesis 3

"Personality, not role": who does it help — and who doesn't it?

We used to write "you are a marketing expert." That's a role: it tells the model who to be, not how to behave. Now OpenAI suggests describing character — e.g. playful, direct, curious. Plus one more parameter — decisiveness: you can literally write "if the task is at least 80% clear, don't ask — just do."

Character is a genuinely useful layer on top of the role. Decisiveness, though, depends on the task. I build my assistants the opposite way — to ask questions, collect full context and never guess: the entire interview method is built on exactly that.

Decisiveness ("80% clear — go"):
+ beginners and people outside their expertise —
  fewer paralyzing questions, faster first result

Questions ("collect context, then act"):
+ experts and tasks built on unique experience —
  the model asks for what's in your head instead of guessing

My honest verdict on the guide: they built something, but why? 😂 Half the advice is great for simple tasks and harmful for expert ones. The only universal filter is knowing which type yours is.

05 — The test

Does your task need a short or a long prompt?

Take the task you do with AI most often and answer four questions:

06 — The verdict

What to take from the guide — and what to leave to OpenAI?

My filter across all three theses: