10 ChatGPT Prompts We Tested for 90 Days — Real Results

In early January we decided to stop guessing and start measuring. The PromptVault editorial team committed to a 90-day experiment: every workday, we would log exactly which ChatGPT prompts we used, how long each interaction took, and whether the output was actually useful or just plausible-sounding text we ended up rewriting anyway. What followed surprised us — and killed several myths we had believed ourselves.

Before we share the results, a caveat: we are a small team running a content and community site. Our workflows involve writing, research, light coding, email, and community moderation. If you work in a radically different field, your mileage will vary. That said, the patterns we found are almost certainly transferable.

Why Most Productivity Advice About AI Is Wrong

The standard advice goes like this: "Use AI to speed up your work — draft emails, summarise documents, write social posts." We tried all of it. And here is the uncomfortable truth: speed is the wrong metric. We were measuring the wrong thing for the first six weeks of the experiment.

The prompts that generated output fastest were almost never the prompts that saved us time overall. We would get a fast draft, spend 20 minutes editing it into something usable, and end up slower than if we had just written it ourselves. The real productivity gain came from prompts that helped us think more clearly, not prompts that put words on a page faster.

That insight reframed the entire second half of our experiment. We started looking for prompts that improved decision quality and reduced rework, not prompts that maximised word count per minute.

The 3 Prompts That Failed (And Why)

We believe in honesty, so let us start with what did not work.

Fail #1: "Write a complete article about [topic]"

Output was always competent and always hollow. The articles read like they were written by someone who had read about the topic but never actually engaged with it — because that is exactly what happened. We junked every single long-form draft the model produced without heavy editing, and the editing took longer than writing from scratch.

Fail #2: "Summarise this document and pull out action items"

Sounds useful in theory. In practice, the "action items" were generic and the summaries missed the nuanced points that actually mattered. The model would confidently surface the obvious conclusions while burying the subtle one that actually required attention. We learned that AI summaries require you to already know the document well enough to spot what it missed — at which point you barely needed the summary.

Fail #3: "Help me reply to this email"

This one cost us time, not saved it. The drafts were too formal, too long, and required so much editing that we would have been faster typing 40 words ourselves. Email, we concluded, is the worst use case for most people — a conclusion backed by our later tracking data (see our post on AI time savings across business workflows).

The 10 Prompts That Actually Worked

1. The Goal-Clarification Prompt (The Counterintuitive Winner)

This is the prompt we least expected to top the list, and it slows you down at first — which is exactly why it works:

"I am about to work on [task]. Before I start, ask me 5 questions that would help me clarify exactly what I am trying to achieve, who this is for, and what success looks like. Do not offer suggestions yet — just ask questions."

Spending 10 minutes answering those questions consistently saved 45–90 minutes of rework. We used this before every major piece of content and before every significant decision. It felt inefficient. It was the opposite. Browse our ChatGPT prompt library for more clarity-first prompts.

2. The Devil's Advocate Prompt

"Here is my plan: [plan]. Argue strongly against it. Find every flaw, assumption, and risk I might be ignoring. Do not balance this — be one-sided against my position."

This caught three significant mistakes over the 90 days, each of which would have cost real time to fix.

3. The First Principles Decomposition Prompt

"I need to [accomplish goal]. Break this down to its first principles — what are the core components, what must be true for this to work, and what is the most common mistaken assumption people make about tasks like this?"

4. The Research Synthesis Prompt

"Here are 5 sources I found on [topic]: [paste excerpts]. Identify the points where these sources agree, where they disagree, and what important question none of them seem to answer."

This replaced about 2 hours of manual comparison work per week for our research tasks.

5. The Structured Outline Prompt

"Create a detailed outline for [piece of content]. Each section should have a clear argument or insight, not just a topic header. Include the one thing most articles on this topic get wrong."

6. The Code Explainer Prompt

"Explain this code to me as if I wrote it 6 months ago and forgot how it works. Focus on the non-obvious decisions and flag anything that might cause problems later."

7. The Meeting Prep Prompt

"I have a meeting about [topic] with [stakeholder type]. Give me the 5 questions they are most likely to ask that I might not be prepared for, and suggest a one-sentence answer to each."

8. The Feedback Translator Prompt

"Here is feedback I received: [feedback]. What is the underlying concern behind this feedback? What would fully addressing it look like in practice?"

9. The Task Breakdown Prompt

"I keep procrastinating on [task]. Break it into the smallest possible concrete steps — each step should take no more than 15 minutes and have a clear physical output."

10. The Learning Accelerator Prompt

"I am learning [skill/topic]. I have been doing it for [time]. What are the 3 things I am probably doing wrong at this stage that most intermediate learners get stuck on? Be specific to my level."

What We Would Do Differently

If we were running this experiment again, we would track prompt iteration count — how many times we had to rephrase a prompt before getting something useful. That number turned out to be a better predictor of time saved than the prompt category. We also learned the hard way that longer prompts are almost always better. The effort of writing a detailed prompt is almost always returned threefold in output quality.

For a deeper dive into why prompt structure matters so much, read our guide on prompt engineering for beginners. And if you want to explore community-submitted prompts that have been tested by real users, search our prompt library.

The Bottom Line

After 90 days, the honest summary is this: ChatGPT is a genuinely powerful tool for thinking, not primarily for writing. The prompts that treated it as a thinking partner — asking questions, stress-testing ideas, synthesising information — consistently outperformed the prompts that treated it as a content generator. If your AI workflow is mostly about getting words on a page faster, you are almost certainly leaving the bigger gains on the table.