Best ChatGPT Prompts for Productivity: 10 Prompts, Organized by What Goes Wrong

Most prompt lists tell you what to type. This one starts from what is going wrong in your current prompts, then gives you the rewrite that fixes it.

By Tom Morgan · Updated October 22, 2026 · 14-minute read · 10 copy-paste prompts

The short version

• The research supports using the tool, not these templates. In the best-known controlled study, competent professionals using ChatGPT finished writing tasks roughly 40% faster, with work graded about 18% higher. Those gains come from using AI at all. No study has tested the templates below, so treat them as hypotheses to test (there’s a five-day method at the end), not guaranteed fixes.
• The same research has a warning attached. On tasks outside what the model handles well, consultants using AI were less likely to reach the correct answer than those working without it. Two of the 10 prompts below exist to catch that.
• One piece of older advice has aged badly. If you use a reasoning model, telling it to “think step by step” is mostly wasted. Ask for assumptions and alternatives instead. Template 1 does that.
• You can check all of this yourself. The last section is a five-day test that takes about two minutes a day and tells you whether a template beats what you do now.

Start here: the 10 prompts at a glance

Skip ahead if you already know which problem you have. The research and the reasoning behind each prompt follow below.

Prompts 1, 8, 9 and 10 work on almost any task; the rest are tuned to specific jobs.

What changed since 2025

The first version was written around GPT-4o. This one works across models, fixes a statistic that merged two studies, updates the chain-of-thought advice, and adds a sixth failure mode (trusting output where the model is weak).

What the research actually shows

Two studies get quoted in nearly every “ChatGPT makes you more productive” post. They are often blended. They measure different things.

Study 1: writing tasks (MIT, Noy and Zhang, published in Science, 2023). Around 450 experienced professionals (marketers, grant writers, analysts, HR staff, managers) completed realistic business writing tasks, some with ChatGPT and some without. Average time per task fell by roughly 40%, and graders who didn’t know which documents were AI-assisted scored the output about 18% higher. Nielsen Norman Group’s write-up re-expressed the same data as 59% more documents per hour. That 59% is NN/g’s recalculation, not a headline figure from the paper, so cite the Science paper when accuracy matters. The researchers’ own reading is that people spent less time producing a first draft and more time editing.

Study 2: consulting tasks (Harvard Business School, Wharton, MIT and BCG, Dell’Acqua and colleagues, 2023). 758 BCG consultants were split into groups. On tasks within the model’s capability, those using GPT-4 completed more work, finished faster, and produced noticeably higher-rated output. On a deliberately chosen task outside that capability, AI users were about 19 percentage points less likely to get the right answer than the control group. The authors call this boundary the “jagged frontier.”

Neither study compared prompt templates against free-form prompting. Both describe what happens when competent people use the tool on real work. The templates in this guide are built on a narrower, practical claim: most wasted time comes from a handful of repeatable prompt mistakes, and each one has a fix you can test in a week.

Failure 1: You ask for the deliverable when you need the reasoning

“Write me a [thing]” gets you a [thing]: fluent, competent, and impossible to defend. The failure is quiet because the output looks finished. You find out it’s shallow when someone asks why and you don’t know.

The 2026 fix. Don’t ask the model to narrate its process. Ask for the parts of the reasoning a decision actually depends on: the assumptions, how likely each is to be wrong, and what changes if it is. That works on reasoning and non-reasoning models alike, and it surfaces something you can check.

Template 1: Assumption audit

text

Task: [your task]Before you answer, list the 3–5 assumptions your answer depends on.For each one, tell me how likely it is to be wrong and what wouldchange in your recommendation if it were. Then give your recommendation.

Where it earns its place: decision memos, risk reviews, anything you’ll have to explain to someone else. Follow it with: “Which assumption is weakest, and what’s the cheapest way for me to test it?” That one follow-up often replaces an hour of scenario planning.

When “think step by step” still helps, and why this note will date. Vendors keep changing how much reasoning happens by default, so check the current guidance for your model rather than trusting any fixed rule, including this one. As of this update, OpenAI says reasoning models work through problems internally, while fast modes or low reasoning settings still benefit from asking the model to outline its approach. The assumption audit is the durable choice: it works regardless of mode, and it gives you something to check.

Failure 2: No role, no audience, no calibration

Without an audience, the model writes for everyone, which means it writes for no one. The fix is specificity about who is reading and what they must decide, not a vague “act like an expert.”

Here is an illustrative before and after (not a client case, just a pattern that is easy to reproduce):

Template 2: Role, audience and decision

text

You are [specific role]. Explain [topic] to [audience: who they are,what they already know, what they care about].They need to decide [decision]. Give them only what that decisionrequires, in under [N] words. Skip any jargon they wouldn't already use.

Template 3: Code review (for developers)

text

Act as a senior engineer reviewing this code. Find the root cause of[symptom], not just the symptom. Propose the smallest change that fixesit and tell me what you'd deliberately leave alone. If you needinformation you don't have (versions, logs, config), ask before guessing.[paste code]

Template 4: LinkedIn post from your own material (for content creators)

text

Draft a LinkedIn post for [job title] using only the raw material below:[paste your notes, a result you got, a mistake you made, or a realclient story]. Use Problem-Agitate-Solution, but invent no statisticsor anecdotes; if something is missing, flag it instead of filling it in.Open with the sharpest tension in the material. Under 150 words.Give me two different opening lines to choose from.

This one works because the substance is yours. A prompt that supplies only a topic produces a post any competitor could have written.

Failure 3: You ask for everything in one message

A prompt with five requirements gets an answer that treats each one shallowly. Splitting the work is slower per message and faster overall, because you can correct course after the first piece instead of after the fifth.

The three-step sequence below (Template 5) takes about eight minutes. It replaces a first-pass literature review that usually takes much longer. Use it when you’re trying to understand a contested topic, not when you need a quick summary.

Template 5, step 1: Find what’s contested

text

Summarize the key findings in the content below. Cover only what iscontested or unresolved: conflicting viewpoints and gaps in the data.Skip what is settled.[paste content]

Template 5, step 2: Find the missing questions

text

Based on the gaps you just identified, what are the three questionsthis content doesn't answer? Rank them by how much they matter tosomeone deciding [specific decision].

Template 5, step 3: Stress-test the top gap

text

For the highest-ranked gap, what is the strongest counterargument to themainstream view? What evidence would I need to find to settle it, andwhere would I look?

One limit worth knowing: this only works on material you paste in. If you ask it to analyze a source it hasn’t been given, you’re back to relying on whatever it half-remembers.

Failure 4: No constraints on length or format

By default, models lean toward comprehensive. Comprehensive is rarely what you need when you’re trying to get something done. A constraint isn’t a restriction on quality. It tells the model what quality means for this task.

Template 6: Decision support

text

Recommend one option for [decision]. Format: a one-sentencerecommendation, three bullets with the strongest reasons, and onesentence naming the biggest risk. Under 150 words total. No preamble.

Template 7: Automation workflow (with a worth-it check)

text

I want to automate [task]. Current manual process: [steps, tools, howoften, who's involved]. Tool: [Zapier / Make / n8n / other].First tell me whether this is worth automating: estimate monthly timesaved from the numbers I gave, and flag any step that needs humanjudgment and should stay manual. If it is worth it, give the trigger,the actions in order, the two likeliest failure points with how I'dnotice each, and what to test before switching it on. If it isn'tworth it, say so and stop.

Letting the model say “don’t automate this” is what separates a useful workflow prompt from a generic one.

Failure 5: You treat the first output as final

Treat the first answer as a draft whose main job is to show you what you actually wanted. People rarely do this, partly because of anchoring. Tversky and Kahneman described the effect in 1974: the first number or framing you see pulls later judgments toward it. A polished AI answer is an unusually strong anchor, because it looks authoritative. So people edit around it instead of asking for a different one.

Template 8: Controlled revision

text

The last response was close. Keep: [what worked]. Change: [what didn't].New requirement: [precise specification]. Revise the existing textrather than starting over.

Template 9: Adversarial check

text

Now argue against the recommendation you just gave. What is the strongestcase that it's wrong? What evidence would invalidate it? Make theargument as strong as you can; don't soften it.

Template 9 is the most underused prompt in professional settings. You have a plan, you believe in it, and for three minutes you ask a model to find the holes. Sometimes it finds nothing. Sometimes it finds the one that would have cost you a quarter.

Failure 6: You trust it where it’s weakest

This is the failure the BCG study put on the map. The model was strong inside its frontier, and a liability outside it, and the users couldn’t tell which side of the line they were on, because the answers sounded equally confident both times.

You can’t see the frontier, but you can make the model show you where its claims come from.

Template 10: Claim audit

text

List every factual claim, number, name, and citation in your last answer.Mark each one as: (a) taken from the material I gave you, (b) from yourgeneral knowledge, or (c) uncertain. For every (b) and (c), tell mehow I can verify it. Do not invent sources.

Run it before anything leaves your hands: client emails, board slides, published copy. It won’t catch every error, and a confidently wrong claim can still be labelled (b). But it turns “I think this is right” into a short checklist of things to confirm, which is the habit the research says people need.

Handling confidential data. Keep sensitive material out of consumer accounts. Samsung restricted employee use of generative AI in 2023 after engineers pasted internal source code into ChatGPT, a data-handling failure rather than a model failure. If the work involves client or proprietary data, use a business or enterprise tier whose data terms your company has actually reviewed, and anonymize before you paste.

For individual contributors: start with one task

The six failures map onto work you probably do weekly: planning, email, research, decision memos, meeting prep.

What to do: choose one recurring task that eats half an hour or more. Build a template for it from the list above, then run the test in the next section. One task, one template, one week.

What will get in your way: a single bad output. One bad output isn’t a verdict on a template, because single results are noisy. Judge it on the week’s average.

Stop doing this: pasting AI output into work you can’t paraphrase. If you can’t explain what it says without rereading it, you don’t own the work yet.

For team leads: a shared prompt library beats individual tricks

An individual’s good prompt saves that individual time. A tested prompt shared across twelve people saves it twelve times, and the person who wrote it learns from the eleven who edit it.

What to build: a short library in whatever your team already uses (Notion, Confluence, a shared doc). Keep it to the 5 to 10 prompts people actually use. For each one, record:

What will get in your way: nobody maintains it unless maintenance is part of one person’s job.

Stop doing this: asking “is the output good?” The useful question is “how much faster did we finish at equal or better quality?” Only the second one supports a business case.

When these templates still produce garbage

A template improves a prompt. It can’t fix the situation around it. Expect poor results when:

• You haven’t supplied the facts. Template 5 only analyzes what you paste in. Without source material, the model fills gaps from memory, and the output reads just as confidently.
• The task needs your judgment, not its drafting. Strategy calls, sensitive conversations, and anything where context lives in your head produce polished, generic answers however well you prompt.
• A tight constraint removes what you needed. A 100-word summary of a complicated decision drops the trade-offs. Loosen the limit rather than blaming the model.
• The role is decoration. “Act as a world-class expert” adds nothing. Specific audience and decision details do the work.
• It’s the model’s weak area. Precise arithmetic on long data, recent events, niche citations, and legal or medical specifics are where Template 10 matters most and still won’t be enough. Verify these independently.

If a template fails twice on the same task, stop tweaking the wording and ask whether the task suits AI at all.

Test it yourself: a five-day check

Published studies tell you what happened to other people on other tasks. This tells you what happens to you. It takes a couple of minutes a day and needs no tools beyond a notes app.

1. Pick one recurring task (for example, the weekly status update or the first draft of a client email).
2. Days 1 and 2: do it your usual way with ChatGPT. Note the three numbers below.
3. Days 3 to 5: use one template from this guide, adapted to the task. Note the same three numbers.
4. Compare averages.

Five data points is a sanity check, not a study. Treat a large, consistent gap as a good sign and a muddy one as a reason to adjust the template, not abandon the approach. If you do run it, I’d like to hear what you found.

Frequently asked questions

Do prompt templates still matter on the newest models?

Less than they used to for simple tasks, and about as much as ever for tasks where the stakes, the audience, or the format matter. Better models forgive vague prompts, so bad prompts look less bad. They don’t tell you what you meant, who is reading, or what you plan to do with the answer, and that context is what the templates supply.

Should I still tell ChatGPT to “think step by step”?

On full reasoning models, generally no. OpenAI’s guidance says they reason internally and that scripting the process is unnecessary. On fast modes or with reasoning set low, asking the model to outline its approach can still help. In both cases, asking for assumptions (Template 1) gives you something you can check.

What’s the difference between custom instructions and a prompt?

Custom instructions and project-level instructions apply to every conversation: your role, tone, format defaults. A prompt applies to one task. Put what never changes in the instructions, and what changes in the prompt. If you work in a tool without that option, open the session with a short setup message that does the same job.

Is it safe to paste work documents into ChatGPT?

Not into a personal account. For anything confidential, use a business or enterprise plan with data terms your organization has reviewed, or a self-hosted model. Where that isn’t possible, anonymize names, figures and identifiers before pasting.

Sources and notes

• Noy, S. and Zhang, W. (2023). “Experimental evidence on the productivity effects of generative artificial intelligence.” Science, 381(6654). science.org/doi/10.1126/science.adh2586
• Nielsen, J. “ChatGPT Lifts Business Professionals’ Productivity and Improves Work Quality,” Nielsen Norman Group. nngroup.com/articles/chatgpt-productivity
• Dell’Acqua, F. et al. (2023). “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality.” Harvard Business School Working Paper 24-013.
• Tversky, A. and Kahneman, D. (1974). “Judgment under Uncertainty: Heuristics and Biases.” Science, 185(4157).
• OpenAI developer documentation: latest-model guide and reasoning best practices. developers.openai.com/api/docs/guides/latest-model
• Samsung’s 2023 generative-AI restrictions after an internal code leak, as widely reported in May 2023.

Update log. October 22, 2026: Corrected the study attribution (the 59% figure comes from the MIT writing study as summarized by Nielsen Norman Group, not from a prompt-template experiment, and the 758-person figure belongs to the BCG study). Updated the chain-of-thought guidance for reasoning models. Added Failure 6, Templates 9 and 10, and a section on when templates fail. Removed an unverifiable anecdote and an unsourced multiplier. Reduced the count to 10 distinct prompts (the research sequence is one prompt in three steps). April 2025: original publication.

Related on BestPrompt.art: Prompt engineering guides · AI productivity templates