ProAI Hub
All articlesGuides

The Best AI Prompts in September 2026: 15 That Actually Work

Prompting changed more this year than in the three before it. Here's what still works, what quietly stopped, and fifteen prompts to steal.

Jeremy Updated September 2026 12 min read

Search "best AI prompts" and you get lists of 500 prompts that all begin with "Act as an expert…". Most of them were written for 2023 models. They're still circulating because nobody ever deleted them.

The models moved on. OpenAI's developer team published guidance this month telling people to clean out their old instructions, because Astra reads nuance well enough that too much detail can now hurt. Anthropic's docs tell you to drop the ALL-CAPS shouting for a similar reason. The advice on most blogs and the advice from the labs themselves are drifting apart.

So this isn't a mega-list. It's fifteen prompts, grouped by the job they do, and a short section on what changed. They work on Claude, ChatGPT and Gemini with small tweaks. If you only read one part, read the next one.

15

copy-paste prompts, grouped by job

up to 30%

better answers on long documents when the question comes last (Anthropic's own tests)

3–5

varied examples Anthropic recommends when format or tone matters

150–300

words: one practitioner guide's rule of thumb for a working prompt

What changed in prompting this year

Prompting didn't die. It split in two. Everyday prompting got easier, because the models read intent better. The serious kind turned into what people now call context engineering: deciding everything the model gets to see, instead of polishing one sentence. Four things follow from that.

Say less, and say why

Anthropic's docs offer the best test for a prompt I've come across: show it to a colleague who knows nothing about your project. If they'd be confused, the model will be too. Their other tip is underrated. Give the reason behind a rule. "Never use ellipses" is a rule. "This will be read aloud by a text-to-speech engine that can't pronounce ellipses" is a reason, and the model can generalise from it.

OpenAI's September 11 guidance for GPT-6 Astra goes further. It was written for coding agents, but the line travels: models handle nuance and ambiguity much better now, so overly specific instructions can hinder results where they used to help. One practitioner guide puts the sweet spot at 150 to 300 words for most tasks. Treat that as a rule of thumb, not a law. The workflow behind it is the useful part: write the shortest version, look at what's actually wrong with the output, and add only the line that fixes that.

Stop shouting

CRITICAL. YOU MUST. NEVER EVER. That style made sense when models ignored you. Now they overreact. Anthropic's docs tell you to dial it back, because recent Claude models respond strongly to the system prompt and can overtrigger. OpenAI's Astra guidance warns that firm boundaries written for older models can make Astra stop where you'd be happy for it to keep going. If your saved prompts are full of capital letters, delete those first.

Long input first, question last, and say what "done" means

Two small habits with an outsized effect. On long inputs, put the documents at the top and your question at the end. Anthropic says queries at the end can improve response quality by up to 30% in its tests, especially with multi-document inputs, and Gemini guides give the same advice. Second, tell the model what finished looks like. OpenAI notes that Astra can be more tentative about how far to take a task and may hand you a first draft and stop. "Done means it runs, you've checked the result, and you've fixed whatever failed" beats hoping.

Context is the real prompt, and checking is the real job

Instead of crafting the perfect sentence, hand the model your situation: who you are, who the output is for, what good looks like. Build that once as a context document and reuse it in a Project or in custom instructions. Prompt 12 builds it for you.

The flip side: production got cheap, checking didn't. A model can write five pages in seconds, but you still have to establish that they're right. That's why several prompts below ask for sources, a confidence level or a self-check, and why prompt 15 exists purely to audit another AI's work. I wrote about the same problem in the GPT-6 Astra breakdown.

One caveat before the prompts. Not every problem is a prompt problem. Anthropic says so itself: before you tune wording, define what success looks like and how you'll test it, and remember that some issues, latency and cost for example, are often easier to fix by choosing a different model. If the model is missing facts, give it the documents.

How each assistant likes to be prompted

Same job, three different tastes. Keep one base prompt and add a small tweak per model. (Still choosing which one to open? Start with the five-assistant comparison or the full LLM comparison matrix.)

Claude

Best for

Long documents and detailed briefs.

Use cases
  • Wrap the parts in tags like <context>, <document> and <task>
  • Give 3 to 5 varied examples inside <example> tags when format or tone matters
  • Put long documents first and your question last
  • Explain the reason behind a constraint, and say what to do instead of what not to do
Pro tip

Skip the CAPS. On Claude Opus 5 you can also drop "double-check your work". Anthropic says it verifies well on its own and extra checks can waste tokens.

ChatGPT

Best for

Strict formats and long, multi-step work.

Use cases
  • Put the output format before the long context
  • Number the steps when order matters, and add validation rules ("valid JSON only, max 3 risks")
  • Define what "done" looks like before it starts
  • Remove strong "always ask first" language written for older models
Pro tip

On OpenAI's o-series reasoning models, "think step by step" may not help and can sometimes hinder, per OpenAI's docs. Test it before you keep it in a template.

Gemini

Best for

Short, scoped, multimodal prompts.

Use cases
  • Keep prompts short and explicit about the output
  • Label every input ("this is the transcript, this is the screenshot")
  • Put your question after the data
  • Add a few examples and ask for an explicit verification step
Pro tip

With very long contexts, placement matters. Don't bury the key instruction in the middle.

The template behind almost every prompt below

Most prompts below follow one skeleton: goal, context, input, what good looks like, what to avoid, format, and what "done" means. You don't need every line every time. For a quick question, one sentence is fine. Use the whole thing when the result matters. A role line is optional. Anthropic says even one sentence can focus tone, but one guide finds it has little effect on factual or classification tasks, so don't lean on it.

Use it when: the result matters and you want a reusable starting point.

Goal: [the outcome I want, and how I'll judge it]
Context: [my situation, who the output is for, what I've already tried, and why this matters]
Input:
---
[paste the material here]
---
Aim for: [what a good answer looks like]
Avoid: [the default bad answer: generic, hedged, padded, flattering]
Format: [exactly how to structure the answer]
Done means: [what "finished" looks like]
If anything essential is missing, ask me up to 3 questions first.
At the end, tell me where you're least sure.

Long input? Put it first and the question last. Using Claude? Wrap the parts in tags like <context> and <task>.

Think and decide

Don't ask a model "is this a good idea?". It's built to be agreeable and you'll get a polite yes. Make it work against you instead.

01. The Red Team

This one asks how the plan dies, not whether it's good.

Use it when: you're about to commit money, time or reputation.

Goal: Find how this decision fails before reality does. Be my adversary, not my cheerleader.
Decision: [2-3 sentences]
Context: [numbers, constraints, deadline, who is affected]
My reasoning: [why I lean this way]
Do this:
1. Restate my decision in one sentence so I know you understood it.
2. Give the 3 most likely ways it fails, ranked by likelihood x impact.
3. Name the assumption I'm leaning on hardest that I haven't tested.
4. Suggest the cheapest test that could confirm or kill it within a week.
5. Verdict: go, go with changes, or stop. One sentence why.
If something essential is missing, ask me up to 3 questions first.

02. The Pre-Mortem

Borrowed from psychologist Gary Klein. The model just makes it fast.

Use it when: you're starting a project and want the failure modes on the table on day one.

It is six months from now and the project below failed badly.
Project: [what it is, the goal]
Team, budget, timeline: [the basics]
Write the post-mortem as if it already happened:
- The 5 most likely causes of failure, each with the early warning sign I'd have seen in weeks 1 to 4.
- Which causes are in my control and which aren't.
- For each controllable cause, one action I can take this week.
Stay specific to my situation. If a cause would apply to any project, cut it.

03. The Two-Sided Case

Use it when: you're torn between two options and your head keeps going in circles.

I'm torn between two options.
Option A: [description]
Option B: [description]
What matters most to me, in order: [criteria]
Step 1: Make the strongest honest case for A, then for B, as each option's best advocate.
Step 2: Tell me what information would change the answer, and which of it I can get within 48 hours.
Step 3: Given my criteria, recommend one. Say how confident you are (low, medium or high) and why.
If your answer is "it depends", say exactly what it depends on.

Write and communicate

The fastest way to sound like an AI is to let it write in its own voice. Give it samples of yours. "Professional yet friendly" produces mush.

04. Write Like Me

Use it when: your drafts keep coming back sounding like nobody. Run part 1 once, then reuse the style guide.

Part 1. I'm going to paste 3 samples of my writing. Analyse them: sentence length, rhythm, vocabulary, how I open and close, and what I never do. Give me a style guide in 8 bullets. Then wait.
Samples:
---
[sample 1]
---
[sample 2]
---
[sample 3]
---
Part 2 (once you have the style guide). Write [the piece] in that style. Skip cliches and stock openers such as "in today's fast-paced world". After the draft, list any sentence you're unsure sounds like me.

05. The Ruthless Editor

Use it when: you have a draft that's fine but flabby.

Edit my draft for [audience]. Goal: shorter and sharper, with my voice and my claims intact.
Rules:
- Cut at least 20% of the words.
- Flag every sentence that could appear in any article on this topic. Replace it with something specific or cut it.
- Flag any claim that needs a source or a number.
- Keep my ideas as they are and keep the tone as it is. Don't make it sound more "professional".
Output: (1) the edited draft, (2) the 5 biggest changes and why, (3) 3 places where I'm avoiding the point.
Draft:
---
[paste]
---

06. The Hard Message

Use it when: you have to send something difficult and you'd rather not sound angry or spineless.

I need to send a difficult message.
To: [person, relationship, power dynamic]
What happened: [facts only]
What I want to happen after they read it: [outcome]
What I want to avoid: [drama, defensiveness, damage to the relationship]
Write 3 versions: direct and short / warm and collaborative / firm with a clear consequence. Under 120 words each.
Then tell me which one you'd send, and the one sentence in it most likely to backfire.

Research and learn

Two rules: make the model show its sources, and make it test you instead of flattering you.

07. The Sourced Research Brief

Use it when: the answer will end up in a deck, a decision or an article. Use an assistant with web search turned on.

Research question: [the question]
What I'll use this for: [decision, article, deck]
Success looks like: [what a complete answer contains]
Prefer information from: [timeframe]
Rules:
- Cite a link or exact title for every factual claim, and check important claims across more than one source.
- Separate what is well established from what is disputed or thinly sourced.
- If sources conflict, show both and say which you trust more and why.
- Flag anything you can't verify instead of filling the gap with plausible-sounding text.
Output: a 1-paragraph answer, then a table (Claim | Source | Confidence: high, medium or low), then "What I would check next".

08. The Gap Finder

Use it when: you think you understand something and want to find out if you actually do.

I want to really understand [topic]. My level: [beginner, intermediate...].
1. Explain it in 5 sentences a smart 12-year-old would follow.
2. Explain it again for someone in [my field], with an analogy from [my field].
3. Ask me 5 questions, one at a time, that test understanding rather than memory. Wait for my answer before the next one.
4. After each answer, tell me what I got right, what I got wrong, and what I'm confusing it with.
At the end, give me a 3-line summary of my gaps.

09. The Two-Week Sprint

Use it when: you want to learn a skill by making something, not by collecting bookmarks.

Goal: learn [skill] well enough to [concrete outcome] by [date].
Time available: [minutes per day]. Starting point: [what I already know].
Build a 14-day plan. Each day: one 30 to 45 minute task, one thing I produce (not just read), and one way to check I learned it.
Put the 3 most common beginner mistakes in this skill at the top.
End with a "proof of work" project I can publish or show.
Keep it concrete. Skip advice like "practise daily".

Work and productivity

The boring prompts that pay for themselves every week.

10. Meeting Notes to Actions

Use it when: you have raw notes or a transcript and need clarity in two minutes.

Below are my raw notes or transcript from a meeting. Output exactly:
1. Decisions made (one line each, with who decided).
2. Action items as a table: Owner | Task | Deadline | Source line. If no owner or date was stated, write UNASSIGNED rather than guessing.
3. Open questions nobody answered.
4. Risks or disagreements that were glossed over.
5. A follow-up email under 120 words I can send today.
Notes:
---
[paste]
---

11. The Honest Weekly Review

Use it when: it's Friday and you suspect your week didn't match your goals.

Here is my week: calendar, tasks done, tasks that slipped, and what's on my mind.
---
[paste]
---
My goals this quarter: [2 or 3 goals]
1. Tell me in 3 sentences where my time went versus my goals.
2. Identify the one thing I keep postponing and the likely reason.
3. Propose next week's top 3 priorities. Everything else goes on a "not this week" list.
4. Tell me what to stop, delegate or shrink.
Be direct. If my week doesn't match my goals, say so.

12. The Context Builder

This is the highest-leverage prompt in the article. Run it once, then paste the result into a Project or your custom instructions.

Use it when: you're tired of re-explaining yourself.

I want to build a reusable context document so you understand my work without me re-explaining it every time.
Interview me. Ask one question at a time about: my role, who my audience or users are, what "good" looks like in my work, my constraints, my tone, the terms I use, and the mistakes AI tools usually make for me.
After 10 questions, write the document in under 500 words, in a form I can paste into a Project or custom instructions. Include a section called "Never do this".
Start with question 1.

Build and automate

For Claude Code, Codex, Lovable, Astra or any tool that turns words into software. Think before you build.

13. Spec Before Code

Skipping this is the most expensive shortcut in vibe coding.

Use it when: you're about to ask an AI to build something.

I want to build: [what, in plain language].
Users and the one job it must do: [description]
Done means: [what "working" looks like, e.g. it runs, I've tested the main flow, failures are fixed]
Don't write code yet. First:
1. Restate what I'm asking in 3 sentences.
2. List the assumptions you're making. Ask me about any that would change the design.
3. Propose the smallest version that proves the idea (max 5 features), and list what you're leaving out on purpose.
4. Describe the data it stores and the screens it needs.
5. List the 5 ways it could break or be abused.
Wait for my approval before writing code.

14. Debug From the Error

Use it when: something is broken and "fix it" isn't working.

Something is broken.
Expected: [what should happen]
Actual: [what happens instead]
Exact error message or logs:
---
[paste]
---
Relevant code:
---
[paste]
---
What I already tried: [list]
Give me: (1) the most likely cause and how to confirm it in under 5 minutes, (2) two other possible causes, ranked, (3) the smallest fix, (4) how to keep it from coming back.
Change only what's needed instead of rewriting my whole file. If you need more information, tell me exactly what to paste.

15. The Verifier

Use it when: an AI or an agent just handed you something you're about to rely on. Run it in a fresh conversation, ideally on a different model than the one that wrote the output.

You are the reviewer, not the author. Below is a task I gave an AI and the output it produced.
Task: [what I asked for]
Output:
---
[paste]
---
Check it like a skeptical colleague:
1. List every factual claim, number and name in the output. Mark each: verifiable from the material I gave / needs an outside source / possibly invented.
2. Find where the output silently changed or ignored my instructions.
3. Find anything that sounds confident but rests on an assumption.
4. Tell me the 3 things I should verify by hand before using this, in order of risk.
Review only. Leave the output as it is and skip the reassurance.

Habits I'd drop in 2026

  • Shouting. CAPS, "NEVER", "YOU MUST". Both Anthropic and OpenAI now say to tone it down.
  • Blanket "think step by step". OpenAI's docs say it may not help on its reasoning models and can sometimes hinder. Anthropic notes that a general nudge like "think thoroughly" often beats a hand-written plan. Test it on your task.
  • The 800-word mega-prompt. Start short and add only the line that fixes a specific failure.
  • Leaning on a persona. A role helps tone, but it's a seasoning, not the meal. State the goal and what good looks like.
  • Old "always ask first" boundaries carried over from earlier models. Astra may take them too literally.
  • Trusting the first answer, especially anything with numbers, names or dates.

How many examples to include is still a knob to test. Anthropic suggests 3 to 5 varied ones. One 2026 benchmark on software-engineering tasks found that 1 or 2 often did best and more could hurt, though code review kept improving up to five.

A routine that actually sticks

  1. 1Run prompt 12 once and save the context document in a Project or your custom instructions.
  2. 2Keep five prompts you'll really use. Start with 01, 05, 07, 10 and 15. Save them somewhere you'll open. A Notion page works, and the AI Second Brain is built for exactly that.
  3. 3When the output disappoints, change the input first: more context, one example, a tighter constraint. Only then change the model.
  4. 4Once a month, delete the prompts you stopped using. A short list you trust beats a long one you don't.

Frequently asked questions

Sources and notes: the prompts are original. The guidance behind them comes from the labs' own documentation plus several practitioner guides, which are opinion rather than research. Test the prompts on your own tasks and adapt them.