Best ChatGPT Prompts: Patterns That Work

Most "500 best ChatGPT prompts" lists are worth very little, and it's worth knowing why before you copy one: they optimize for looking impressive rather than for the thing that actually makes a prompt work, which is context the model doesn't otherwise have.

"Act as a world-class marketing expert and write compelling copy" contains no information. It reads like a prompt and performs like a shrug.

Six patterns do the real work — and they line up closely with what the model vendors themselves publish, like Anthropic's prompt engineering guide and Google's prompt engineering guide. Learn these and you can write a better prompt for your situation than any list can supply:

1. Give it the raw material — paste the actual document, data, or transcript
2. Define the output shape — format, length, and structure, explicitly
3. Show one example — a single good sample beats three paragraphs of description
4. Name the reader — who this is for and what they already know
5. Set the constraints — what to avoid, what must appear, what's out of scope
6. Ask for the reasoning when it matters — and for the uncertainty, so you know what to check

Everything below is these six applied. Copy the examples, then stop copying and start composing.

Quick Comparison: Weak vs Strong

The Six Patterns, With Examples

1. Give it the raw material

The single biggest quality jump available. A model reasoning about your actual meeting transcript beats one reasoning about the idea of a meeting.

Here is the transcript of a customer call: [paste]. List every objection raised, quote the sentence it came from, and mark whether we answered it.

The quoting requirement matters — it forces the answer to stay anchored to the source and makes it checkable in ten seconds.

2. Define the output shape

Ambiguity in the ask becomes filler in the answer.

Rewrite this as a table with columns: Risk, Likelihood, Impact, Owner. One row per risk. No preamble, no summary paragraph.

"No preamble" is worth adding to almost everything. It removes the reflexive throat-clearing.

3. Show one example

Faster and more reliable than describing a style.

Here are two headlines in our voice: [example], [example]. Write eight more for this article: [paste]. Match the rhythm and the register, not the topic.

4. Name the reader

Same content, different reader, different piece — and the model can't guess.

Explain this error message to a designer who's never used a terminal. No jargon. Tell them what to do next, not what went wrong internally.

5. Set the constraints

Constraints are where taste lives.

Draft a reply declining this vendor. Two short paragraphs. Warm but final — no "let's revisit next quarter." Don't explain our budget.

6. Ask for reasoning and uncertainty

For anything you'll act on.

Analyse this spreadsheet [paste] and tell me which region underperformed. Show the numbers you used. Say explicitly what you're unsure about and what data would settle it.

That last sentence is the highest-value clause most people never write. It converts a confident answer into a checkable one.

Prompts by Job

Project and work prompts

I'm starting [project]. Here's what I know: [paste]. Ask me the ten questions you'd need answered to plan this properly. Don't plan it yet.

Inverting the direction — model asks, you answer — is consistently underused and consistently good. It surfaces the gaps you didn't know you had.

Here are my notes from four calls this week [paste]. What's the pattern I'm not seeing?

Writing prompts

Here's my draft [paste]. Cut 30% without losing any argument. Show me what you cut and why.

Read this and tell me where a skeptical reader stops believing me.

The second one is better feedback than most human reviewers give, because it's specific about the failure mode.

Analysis prompts

Here's our data [paste]. Give me three hypotheses that would explain this, ranked by how easy each is to test.

Learning prompts

Explain [concept] to me. Then ask me three questions to check whether I actually understood it, and correct my answers.

What Not to Bother With

Being specific is more useful than another list of things to try.

• Elaborate role-play preambles. "You are a Harvard-trained strategist with 30 years of experience" does very little. Context does the work, which is the consistent finding across serious prompt engineering practice.
• Threats and bribes. "This is very important to my career" and offering tips are folklore. Spend the tokens on context instead.
• Prompts promising money. "ChatGPT prompts to make money" is a large search and a poor investment — no prompt substitutes for having something to sell. The prompts that help with a real business are the boring ones above, applied to your actual customers and numbers.
• Ultra-long prompt templates copied from a list. They're tuned to someone else's situation, and the parts you don't understand are the parts that will mislead you.

The Part That Compounds

Here's the thing the prompt lists structurally can't give you: the value isn't the prompt, it's the sequence.

Real work is rarely one prompt. It's a chain — extract the data, check it, reshape it, draft, critique the draft, revise. When you get a good result, the reusable asset is that whole sequence with its context and constraints, not the final message you happened to send.

Almost nobody saves this. The good chain lives in one conversation, and the next time the task comes up it gets rebuilt from memory, worse. Memory features help an individual assistant recall your preferences; they don't hand you the sequence back as something you can run again or give to a colleague.

Two habits that fix most of it:

1. Keep a prompt file. A plain document with your working prompts, grouped by task, with a line on what each is for. Unglamorous and more valuable than any downloaded pack.
2. Save the chain, not the message. Write down the steps in order. That's the thing that turns a good afternoon into a repeatable procedure.

If you'd rather that sequence become something you press a button to re-run — and something a colleague can run too — that's the gap Taku is built for: mirror an AI setup someone already proved out, run it in your own desktop workspace, and keep it. Taku is in Beta, and the Mac app is available now. Our piece on why most AI tools end up in your bookmarks covers why so little of this sticks, and how do I use AI is the starting point if you're earlier than this.

Key Points

• Context beats phrasing. Pasting the real material outperforms any clever wording
• Six patterns cover almost everything: raw material, output shape, one example, named reader, constraints, and requested uncertainty
• "No preamble" and "say what you're unsure about" are the two highest-value clauses most people omit
• Skip role-play preambles, threats, and copied mega-templates — they add words, not information
• The reusable asset is the sequence, not the single prompt — write the chain down
• Keep a prompt file. It compounds; a downloaded pack doesn't

FAQ

What are the best prompts for ChatGPT?

The ones carrying your actual context. A prompt that pastes your real document and specifies the output shape, the reader, and the constraints will outperform any generic template. Learn the six patterns and write your own — that's genuinely faster than searching a list.

Do "act as an expert" prompts work?

Barely. Assigning a role adds almost no information, and the model doesn't gain expertise from being told it has some. What changes the answer is giving it your material and telling it who the output is for.

What should I ask ChatGPT to be more productive?

Ask it to interrogate you before it produces anything: "Ask me the ten questions you'd need to plan this properly." Most weak output comes from missing context, and this surfaces it before you've wasted a draft.

Are paid prompt packs worth buying?

Generally no. They're static, they go stale as models change, and they're tuned to someone else's situation. Ten prompts you wrote for your own recurring work will outperform a thousand you bought.

How long should a prompt be?

As long as the necessary context, and no longer. A three-page paste of the actual document plus two sentences of instruction is usually right. Three pages of instruction with no source material is usually wrong.

How do I stop ChatGPT from being generic?

Give it something only you have — your data, your draft, your transcript, your constraints. Generic input produces generic output, and no amount of prompt phrasing fixes an empty prompt.

What's the difference between a prompt and a workflow?

A prompt is one message. A workflow is the ordered sequence of prompts, checks, and edits that gets a real task done. Most people optimize the first and never write down the second, which is why the same work gets rebuilt every time.