Study: Generative AI results depend on user prompts as much as models
Hard evidence that half the 'better AI' effect is the user's prompting skill, the clearest answer to why identical tools produce wildly different results.
Open mitsloan.mit.edu →
Strategy & The AI-Native Founder
Article
Video
Guide
Article
Video
Article
Guide
Article
Article
Article
Video
Guide
Guide
How founders use AI for Prompting
3 questions founders actually ask, each with a straight answer and the resources worth your time.
Why do some people get 10x better results from the same AI tools? #
Because the output you get depends as much on how you ask as on the model itself, an MIT study found that roughly half the quality gains from a better AI model came from users learning to communicate with it, not from the model. Power users treat AI like a smart new hire: they give it background, examples, and clear success criteria, then iterate on the answer instead of accepting the first draft. The good news is this is a learnable communication habit, not a technical skill.
The Skill That Separates AI Power Users From Everyone Else (Why "Clear" Specs Produce Broken Output)
A practitioner breaks down the real gap: instructions that feel clear to you are ambiguous to AI, and how power users close that gap.
Open youtube.com →Using AI Right Now: A Quick Guide
The Wharton professor most founders trust on AI explains which model to use for what and the working habits that compound into better output.
Open oneusefulthing.org →OpenAI report reveals a 6x productivity gap between AI power users and everyone else
Data from OpenAI's own enterprise usage showing the gap is behavioral, not access, everyone has the same tools, few use them intensively.
Open venturebeat.com →What are the 5 prompting techniques worth learning (and the ones that are outdated)? #
The techniques that still pay off: give the AI relevant context and background, show it 1-3 examples of what good output looks like (few-shot), break big tasks into steps, tell it exactly what format you want, and iterate, decompose and refine instead of accepting draft one. Meanwhile, old tricks like elaborate role-play personas ('you are a world-class expert...'), threatening or tipping the model, and manually writing 'think step by step' matter far less with modern reasoning models. In short: clarity, examples, and context beat magic phrases.
AI prompt engineering in 2025: What works and what doesn't | Sander Schulhoff
Directly answers this question, the top 5 techniques that work plus why role prompting and threatening the AI stopped working, from the researcher behind the largest prompting study.
Open youtube.com →Prompt Engineering in 2025: The Latest Best Practices
A practitioner's deep-dive with real examples from AI companies like Bolt, separating research-backed techniques from folklore.
Open news.aakashg.com →Prompt engineering overview (official Claude docs)
The model maker's own playbook, clear, ordered from most to least effective, and kept current as models change.
Open platform.claude.com →We Tested 25 Popular Claude Prompt Techniques: These 5 Actually Work
Rare empirical test of popular techniques against a current model, shows which tips are measurable wins and which are myths.
Open dreamhost.com →How do I give AI the right context about my business (projects, custom instructions, memory)? #
Stop re-explaining your company in every chat: write a one-page brief about your business (what you sell, who your customers are, your tone, your goals) and put it where the AI can always see it. In ChatGPT that means custom instructions plus Projects with your key docs uploaded; in Claude, Projects with a knowledge base works the same way; memory then picks up preferences over time automatically. Once the context lives in the tool instead of in your prompts, every answer starts from 'knows my business' instead of 'generic advice'.
Getting started with Claude Projects: Starting the day off right
A top VC walks through a real founder's Claude Project setup, a concrete, copyable example of putting business context into an AI workspace.
Open sarahtavel.com →The ChatGPT Feature 90% of Users Ignore (But Shouldn't)
A billion-dollar founder shares his exact custom-instruction templates so ChatGPT permanently knows who you are and how you work.
Open simple.ai →
Creating a contextual workspace in ChatGPT
OpenAI's own short walkthrough of using Projects with files and instructions to give every chat standing context.
Open youtube.com →Projects in ChatGPT (official help guide)
The definitive reference on how project files, project-level instructions, and project memory actually behave, worth skimming before you set yours up.
Open help.openai.com →Memory FAQ
Explains what ChatGPT remembers about you, how to view and edit it, and how to keep sensitive business info out, essential hygiene for founders.
Open help.openai.com →