Why prompts fail | Master AI Automation in 4 hours Master AI Automation in 4 hours Course About Ayush Modules Sample chapter Toolbox The Microcap Minute Classroom / Module 02: Talking to Machines: Prompting / Chapter 1 Why prompts fail Watch first, then read. Same lesson, your pace. What you will learn – The single cause behind almost every bad AI answer – What a model does with vague instructions – The two-sentence fix that upgrades any prompt The vagueness tax Try an experiment in your head. A friend says: “Make food.” What do you cook? Breakfast or dinner? For four people or forty? Spicy or bland? You’d have to guess, and whatever you make has maybe a one-in-ten chance of being what they wanted. When you type “write about climate change” or “make it better” or “fix my essay” , the model is that friend in the kitchen. It must guess: length? audience? tone? purpose? language level? It guesses from the most average patterns on the internet, and serves you the most plausible answer instead of the right one. Every guess the machine makes for you is a tax on quality. Call it the vagueness tax . Bad prompts pay it constantly; good prompts refuse to. What the machine actually hears An LLM has no access to your intention, only to your words. So: “Short summary” → short by whose standard? Two lines or two paragraphs? “Professional tone” → banker-professional? teacher-professional? startup-founder-professional? “Better” → better how? Shorter? Funnier? More evidence? The model resolves all of this toward the statistical middle. That’s why bad answers feel so beige . The fix isn’t a magic phrase, it’s supplying the missing specifications yourself, before the machine invents them. The two-sentence fix For any prompt, ask two questions and answer both inside the prompt: “Who is this for, and what should they feel/do afterwards?” “What would I have to tell a human intern to get this right first try?” That second question is the whole trick. Anything you’d need to tell an intern, context, examples, format, limits, what to avoid, the model needs too. If your prompt wouldn’t survive being read aloud to a new intern, it isn’t done. Before: Improve my email After: Here is an email to my teacher asking for a deadline extension (pasted below). Rewrite it: polite but confident, no over-apologising, under 120 words, keep the reason for the delay honest but brief. End with a clear ask. Same tool, same user, different universe of output. Try it yourself Take three prompts you’ve actually typed recently (anything counts). For each: write down every guess the model had to make. Rewrite each with the intern test applied. Re-run old vs new side by side in Kimi or ChatGPT and compare. Save the before/after pair into prompts/ , this comparison is Chapter 2’s launchpad. Key takeaways – Bad outputs are usually missing specifications, not missing intelligence. – Vague prompts push the model to the statistical middle, competent, beige. – The intern test converts vibes into instructions. – Specify audience, purpose, format, length and constraints up front. Download the exercise sheet (PDF) Module workbook (PDF) ← Module 02 index Next: The R-C-T-F recipe → Classroom / Module 02: Talking to Machines: Prompting / Chapter 1 Why prompts fail What you will learn – The single cause behind almost every bad AI answer – What a model does with vague instructions – The two-sentence fix that upgrades any prompt The vagueness tax Try an experiment in your head. A friend says: “Make food.” What do you cook? Breakfast or dinner? For four people or forty? Spicy or bland? You’d have to guess, and whatever you make has maybe a one-in-ten chance of being what they wanted. When you type “write about climate change” or “make it better” or “fix my essay” , the model is that friend in the kitchen. It must guess: length? audience? tone? purpose? language level? It guesses from the most average patterns on the internet, and serves you the most plausible answer instead of the right one.
Why prompts fail
Written by
in