The myth of the perfect prompt
Somewhere out there, people imagine expert prompters typing one flawless incantation and receiving perfection. Nope. Experts just iterate faster. They treat the first output as a draft to react to, not a verdict to accept.
Prompting is a loop:
- Draft a decent prompt (using everything from Lessons 1-8).
- Inspect the output. What's wrong with it, specifically?
- Refine the prompt to fix that specific thing.
- Repeat until it's good, usually 2-3 laps, not 20.
The skill isn't avoiding bad outputs. It's diagnosing them. Every disappointing output is the model telling you exactly which instruction was missing.
Watch a job posting evolve
Let's iterate a real task: writing a job posting for a junior barista. Three laps, start to finish.
Lap 1: the honest first draft
Write a job posting for a junior barista at my coffee shop.
Output (excerpt): "We are seeking a dynamic, passionate coffee professional to join our fast-paced team! The ideal candidate is a self-starter with a growth mindset who thrives in a high-energy environment..."
Inspect: Corporate sludge. Our shop is a scrappy neighborhood place, and this reads like a bank wrote it. Also missing: pay, hours, how to apply.
Lap 2: fix the diagnosed problems
Write a job posting for a junior barista at Marigold Coffee, a small neighborhood shop.
Tone: friendly and down-to-earth. Zero corporate buzzwords - specifically ban "dynamic", "passionate", "fast-paced", and "self-starter".
Include: pay is 18-22 dollars per hour plus tips, weekends required, no experience needed - we train.
Length: under 150 words.
Output (excerpt): "Marigold Coffee is hiring a junior barista. We're a small neighborhood shop where regulars have usual orders and the playlist is a group project. No experience needed. We'll teach you everything from dialing in espresso to latte art. Pay is 18-22 dollars per hour plus tips. Weekend availability required..."
Inspect: So much better. But it buried the pay at the bottom, and there's no call to action. One more lap.
Lap 3: polish the last details
Keep everything about that last version, but: put pay and hours in a short bulleted "Details" section near the top, and end with one line telling people to email hello@marigold.example with the subject "Coffee me".
Inspect: Done. Three laps, maybe four minutes total, and notice lap 3 didn't restate the whole prompt. In a conversation, you can just steer: "keep everything, but change this one thing."
Refine the diagnosis, not the vibes. "Make it better" teaches the model nothing. "Remove the buzzwords, move the pay up top, add a call to action" is a fix the model can actually execute. Vague dissatisfaction in, vague revision out.
The diagnostic cheat sheet
| Symptom in the output | The missing instruction |
|---|---|
| Generic, could-be-anyone content | Context: facts about your situation |
| Wrong tone | A tone spec, or banned-words list |
| Wrong length | An explicit limit |
| Wrong structure | A format spec (bullets, table, sections) |
| Made-up details | The real facts, plus "don't invent details" |
| Right content, wrong emphasis | "Lead with X" or a section order |
โ Weak prompt
Prompt
Follow-up after a disappointing draft: That's not what I wanted. Try again.
Output
A second draft that is different from the first in random ways - and no closer to what you actually wanted, because the model still doesn't know what was wrong.
'Try again' rerolls the dice; it doesn't aim them. The model changed things you liked and kept things you hated, because you never said which was which.
โ Good prompt
Prompt
Follow-up after the same draft: Good structure - keep it. But cut the buzzwords, mention pay (18-22/hr plus tips) in the first three lines, and make the whole thing 30 percent shorter.
Output
A revision that preserves the structure you liked, surfaces the pay early, drops the sludge, and tightens the length - because every change was requested by name.
Praise what works, name what doesn't, give the missing facts. That's a code review, not a coin flip, and it converges in one lap instead of five.
Checkpoint
Prompting is a draft-inspect-refine loop: diagnose specifically what's wrong with each output and fix that named thing, instead of rerolling and hoping.
๐ Quiz
Question 1 of 3What separates expert prompters from beginners, according to this lesson?