The Prompt Mistakes I Made for 6 Months Before Noticing
The frustrating thing about prompt mistakes is that they don't fail loudly. The model doesn't return an error. It produces output. The output looks fine. You ship it, or you edit it for a while, or you give up and rewrite it yourself — and you blame the model for being mediocre when the actual problem is that you asked it the wrong way.
I spent the first six months of using AI seriously making the same handful of mistakes over and over without noticing. The output was always *fine*. None of it was good enough to compound. Once I figured out what I was doing wrong, the same models started producing dramatically better work — same tools, same tasks, just different prompts.
These are the five mistakes that cost me the most.
1. Being vague when I thought I was being clear
For months, my default prompt for any writing task was some version of "Write a blog post about X." Or "Help me write a landing page for Y." I thought I was being concise. I was being lazy.
The problem isn't the length of the prompt. The problem is that "write a blog post about X" doesn't tell the model what kind of blog post, for what audience, with what argument, of what length, in what voice. So the model defaults to the average of all blog posts on the internet — generic structure, hedge-heavy tone, no opinion. The output is technically a blog post. It's also unusable.
The fix is not to write longer prompts. It's to specify the things that actually constrain the output. Audience, argument, length, voice, what to avoid.
This is the structure I now use for any writing prompt:
Write a [LENGTH] [FORMAT] for [AUDIENCE] that argues [ONE-SENTENCE THESIS]. Voice: [DESCRIPTION + EXAMPLE]. Structure: [OUTLINE]. Avoid: [SPECIFIC PATTERNS]. Open with [TYPE OF OPENING]. End with [TYPE OF ENDING].
Category: writing. Tool: Claude. Status: favorite. None of those fields are optional. The first time I filled them all in, the output was so much better than my "write a blog post about X" baseline that I felt stupid for the months I'd spent not doing it.
2. Asking for too many things in one prompt
The second mistake was the opposite of the first. After I learned that more specific prompts produce better output, I overcorrected. I started writing prompts that asked for ten things at once. Write the post, but also suggest three alternative titles, but also generate a meta description, but also write a tweet thread, but also propose three image directions.
What I got back was ten things, all done at 60% quality. The post was decent but not great. The titles were generic. The meta description was bland. None of the ten outputs was as good as it would have been if it had been the only thing the model was trying to do.
Models trade off depth against breadth. Ask for one thing and the model will think about that one thing for the entire response budget. Ask for ten things and the model splits its attention ten ways.
The fix is to chain prompts instead of stacking them. Write the post first. *Then* in a follow-up message, ask for three alternative titles. *Then* the meta description. The total time is barely longer, and every output is dramatically better.
3. Never iterating on prompts that "worked"
This is the mistake I'm most embarrassed about. For months, I'd write a prompt, get an output that was *acceptable*, and use that prompt every time afterwards. If the output was usable, the prompt was done. I didn't iterate.
The problem with "acceptable" prompts is that they leave most of the model's capability on the table. A prompt that gets you a 7/10 output the first time can almost always get you a 9/10 output by version 4 — but only if you treat the prompt as a draft that needs editing the same way you'd treat a piece of writing.
Now, every prompt I use regularly goes through at least three iterations. After each use, I ask myself: what's the one thing about this output that disappointed me? Then I edit the prompt to address that one thing. Run again. Repeat.
This is the meta-prompt I use to iterate:
I've been using this prompt: [PASTE PROMPT]. Here's a recent output it produced: [PASTE OUTPUT]. The thing I dislike most about this output is [SPECIFIC ISSUE]. Suggest three changes I could make to the prompt — not the output — that would address this without breaking the parts that work. Be specific about which lines to change.
Category: prompt-tuning. Tool: Claude. Status: favorite. Most of my favorite prompts are at version 4 or 5, and each iteration was driven by exactly one specific complaint about the previous version's output.
4. Letting the model be agreeable
For a long time, every time I asked a model to review my work, the response was some variation of "this is great, here are a few small suggestions to consider." Useless. Not because the model can't be critical — but because by default, every model is trained to be agreeable in the absence of permission to be otherwise.
The fix is to explicitly authorize disagreement. And to be specific about *what kind* of disagreement you want. "Be critical" isn't enough — the model will produce performative criticism that's softer than it sounds. What works is asking for one specific kind of negative judgment.
This is my critique prompt:
I'll paste a draft / decision / design. Don't tell me what's good. Don't list strengths. Tell me: (1) the single biggest weakness, named specifically — not "could be clearer" but "this paragraph contradicts the previous one." (2) The one thing I'm probably wrong about. (3) The decision a thoughtful critic would push back on hardest. Be direct. No softening. If nothing's wrong, say nothing's wrong — don't invent issues.
Category: review. Tool: Claude. Status: favorite. The "if nothing's wrong, say nothing's wrong" line is what stops the model from manufacturing false criticism just to sound thorough. I added that line after watching it invent problems three sessions in a row.
5. Treating each conversation as a fresh start
The last mistake took me longest to notice. Every time I'd open a new chat, I'd write the prompt from scratch. Even for tasks I'd done dozens of times. I'd remember some of the constraints from last time, forget others, and end up with output that was a step back from what I'd gotten the previous week.
The cost compounds. The hundredth time you write the prompt for a thing you do every week, you're an hour deep into total time spent rewriting the same instructions. And because each version is slightly different, your output quality oscillates instead of compounding.
The fix is to keep prompts. Save them. Name them. Version them. Treat them like small pieces of code you reuse, not like throwaway messages.
This is what changed when I started doing it: my baseline output quality went up *and* stopped fluctuating. I knew exactly what each saved prompt produced. When the output disappointed me, I knew it was the prompt's fault — and I could iterate on the prompt rather than guessing whether I'd phrased it slightly differently this time.
(This is also the reason I built Super Prompts — the moment I started taking prompt reuse seriously, my notes app stopped being a tool and started being a graveyard.)
What the mistakes have in common
If you look at the five mistakes together, they all fail the same way. Each one produces output that's "fine." Each one makes you blame the model when the model isn't the problem. Each one is invisible until you stop and compare what you're getting now against what you'd get if you fixed the prompt.
The reason prompt mistakes go unnoticed for months is that they're not loud. They don't crash. They don't error. They just produce mediocre work, consistently, in a way that feels like the ceiling of what AI can do.
It's not the ceiling. It's the floor. The ceiling is much higher than most people ever reach, because the ceiling lives behind a small number of fixable habits — being specific, asking for one thing at a time, iterating on what works, authorizing disagreement, and never starting from scratch when you don't have to.
If you're using AI every day and the output isn't getting better over time, the problem is almost never the model. It's that you're making one of these five mistakes, every day, and not noticing. Fix them in the order above. The quality of your output will compound from week one.