GPT prompting guide.
Which GPT versions still exist, which were quietly retired, and the six techniques that work across all of them. Start with the version table, because several popular guides describe models you can no longer use.
GPTPrompts.AI Editorial
Model status verified against OpenAI's own retirement announcements · Last updated August 2026
01 · Check this first
Which GPT versions still exist
A lot of prompting advice online is written for a specific version number, and several of the most searched-for versions were retired from ChatGPT earlier this year. If you came here looking for a GPT-4.1 or GPT-5.1 guide, this is the part that matters.
| Version | Status in ChatGPT | Detail |
|---|---|---|
| GPT-5.5 | Available | Released 23 April 2026, in Thinking and Pro variants. Not available on the free tier. |
| GPT-5.4 | Available | Released 5 March 2026. Thinking and Pro first, then mini and nano variants. |
| GPT-5.2 | Available | Released 11 December 2025. Was the model most users had migrated to before 5.4 and 5.5. |
| GPT-5.1 | Retired | Removed from ChatGPT on 11 March 2026, including Instant, Thinking and Pro. |
| GPT-5 (Instant, Thinking) | Retired | Removed from ChatGPT on 13 February 2026. |
| GPT-4.1 and GPT-4.1 mini | Retired | Removed from ChatGPT on 13 February 2026, alongside GPT-4o and o4-mini. |
API availability follows a different schedule from ChatGPT. A model removed from the app can remain callable through the API for a period, and OpenAI publishes those dates separately. If you are building rather than chatting, check the deprecations page linked at the end.
02 · The template
One template that works on every version
Version-specific tricks come and go. This structure has survived every model change because it is not a trick: it is just saying the things the model cannot infer.
You are a [role with relevant expertise]. Your task is to [one clear sentence describing the goal]. Context: [Background, and paste any material the answer should be based on.] Constraints: - [What to avoid] - [Length or depth] - [Tone and audience] - If you are unsure of a fact, say so rather than estimating. Format: [Exact shape of the output, for example: a markdown table with columns X, Y, Z, followed by a two-sentence summary. Nothing else.]
03 · Techniques
Six techniques, each with a line you can paste
1. Specify the format, not just the task
The most common failure in an otherwise good prompt is leaving the output shape unstated. The model has to guess whether you want prose, bullets, a table or code, and a wrong guess makes a correct answer useless. Name the format explicitly, including what should not appear.
Return the answer as a markdown table with exactly three columns: Issue, Severity (High/Medium/Low), Suggested Fix. No introduction, no closing summary, no commentary outside the table.
2. Give it material rather than trusting recall
A model answering from training data is reconstructing, not retrieving. Pasting the actual document, error log or policy changes the task from remembering to reading, which is the single biggest reduction in invented detail you can achieve.
Answer using only the document below. If the document does not contain the answer, say 'not covered in this document' rather than filling the gap from general knowledge. <document> [paste here] </document>
3. Show, do not describe, when the format is fussy
When the output has a stable shape you care about, two or three worked examples beat any amount of description. This is few-shot prompting, and it is most valuable for classification, extraction and anything with a house style.
Rewrite each line in our house style. Follow the examples exactly. Input: We are pleased to announce a new feature. Output: New feature: [name]. Here is what it does. Input: Please be advised that maintenance is scheduled. Output: Maintenance on [date]. Expect [effect]. Input: [your line] Output:
4. Ask for the reasoning to be checkable
Rather than asking the model to think step by step, which reasoning variants already do internally, ask it to surface the parts you would need in order to disagree with it: assumptions, what would change the answer, and confidence.
Before your recommendation, list the three assumptions it rests on, and one piece of information that would change your answer if it turned out to be different. Then give the recommendation in two sentences.
5. Iterate with a diff, not a rewrite
When the answer is close, do not rewrite the prompt. Reply with the single change you want. The model keeps everything else, and you avoid accidentally undoing the parts that already worked.
Good, keep the structure. Two changes: make the tone more direct, and cut the third section entirely.
6. Give it permission to refuse the frame
Models are tuned to be agreeable, which means a flawed premise in your question tends to survive into the answer. Explicitly inviting pushback is a cheap correction.
If the approach I described is the wrong way to solve this, say so first and explain why, before answering the question as asked.
Frequently asked questions
Is there a GPT-5.1 prompting guide?
Not one worth following, because GPT-5.1 was removed from ChatGPT on 11 March 2026, including the Instant, Thinking and Pro variants. Any guide written specifically for it describes behaviour you can no longer reproduce. The techniques transfer though: use the structure on this page against GPT-5.5, 5.4 or 5.2 instead.
Is there a GPT-4.1 prompting guide?
GPT-4.1 and GPT-4.1 mini were retired from ChatGPT on 13 February 2026, along with GPT-4o and o4-mini. If you are searching for GPT-4.1 prompting advice for ChatGPT, the model is gone. API availability for older snapshots has its own timeline, so check OpenAI's deprecations page if you are calling the API rather than using the app.
What is the single most effective change to a GPT prompt?
Specifying the output format. Most weak prompts describe the task well and leave the shape of the answer to chance, so the model guesses and usually guesses wrong. Adding a sentence such as: return a markdown table with columns Name, Risk, Recommendation, and nothing else, fixes more bad outputs than any other single edit.
Do I need different prompts for Thinking and Instant models?
Broadly yes. Reasoning-oriented variants benefit from being given the goal and the constraints and then being left alone, since instructing them to think step by step duplicates what they already do internally. Faster variants benefit from more explicit structure and worked examples. Same prompt content, different amount of scaffolding.
What is a good general-purpose GPT prompt template?
Role, task, context, constraints, format. In practice: You are a [role]. Your task is to [one sentence]. Context: [background and any material]. Constraints: [what to avoid, length, tone]. Format: [exact shape of the output]. Filling those five slots takes a minute and outperforms almost any clever phrasing.
Does prompt engineering still matter as models improve?
The tricks matter less and the specification matters more. Older workarounds like elaborate role-play framing have largely stopped being necessary. Clearly stating what you want, what you have, and what the answer should look like has not become less important, because no model can infer requirements you never said out loud.
How do I stop GPT inventing facts?
You cannot stop it entirely, because generating a plausible continuation is the mechanism rather than a bug. You can reduce it: give it the source material instead of relying on recall, ask for a citation next to each factual claim, and add an explicit instruction to say it does not know rather than estimate. Then check anything specific.
Sources
- OpenAI: retiring GPT-4o and older models
- OpenAI Help Center: model release notes
- OpenAI API deprecations
Model lineups change often. Where this page and OpenAI disagree, OpenAI is right.