Everyone's using AI now. Almost nobody's using it well.
I've spent the better part of two years watching teams adopt ChatGPT and Claude — and I kept seeing the same pattern: people paste a vague question, get a mediocre answer, shrug, and move on. The model isn't broken. The prompt is.
After logging my own usage, I found that five specific prompt mistakes were responsible for almost all of my wasted time and bad outputs. Fix them and you don't just get better answers — you get them faster, which is the whole point.
Here are the five, and the fix for each.
Mistake 1: Asking instead of instructing
The single most common mistake. A prompt like:
"What do you know about caching?"
...is a question. The model will give you a textbook. You didn't want a textbook — you wanted a decision.
The fix — give it a role and an output shape:
You are a senior backend engineer. I run a Node.js API on a single server, 5K req/min, mostly read-heavy product lookups. Recommend a caching strategy. Give me:
1) The one change with the highest impact and why.
2) Two alternatives ranked by effort.
3) The single failure mode I should monitor.
Don't lecture me on what caching is.
Notice: role, constraints, output shape, and an explicit "don't." That last line is worth its weight in gold — it kills the generic preamble.
Mistake 2: Asking for everything, getting nothing usable
"Rewrite my whole codebase to be better" produces 2,000 lines of confident nonsense. The model does better with a focused, ranked ask.
The fix — constrain scope and force ranking:
Review this function. Do NOT rewrite it. Instead:
1) List the top 3 bugs/risks, ranked by severity, with line numbers.
2) For each, give a one-line fix and a one-line reason.
3) End with: "Apply fixes 1-3? Reply Y/N."
The confirmation gate (Reply Y/N) is the trick. It stops the model from making 40 reckless edits and turns it into a co-pilot that waits for you. This single pattern cut my debugging time roughly in half.
Mistake 3: Zero context, then complaining about the output
The model isn't psychic. If you say "write a blog post about Docker," you get a bland post that sounds like every other Docker post. If you give it 30 seconds of context, you get something usable.
The fix — paste a tone reference and an audience:
Write a 600-word intro to Docker containers for a junior dev audience.
Match this tone exactly (it's mine):
"""[paste one paragraph you've already written]"""
Avoid: "In today's fast-paced world..." and any bullet under 3 items.
A tone reference does more than any adjective. "Make it engaging" is meaningless to an LLM. A sample paragraph is an instruction it can actually follow.
Mistake 4: Accepting the first draft
People treat the first response as the deliverable. It's the first draft. The real time savings come from one tight follow-up:
The fix — targeted correction, not a restart:
- "Make section 2 more concrete — give me a real example with numbers."
- "You hedged in the conclusion. Commit to a recommendation."
- "Cut 30% of the words without losing the three insights."
Instead of re-rolling the whole prompt (which gives you a different mediocre draft), you steer the good draft to great. This is where the hours actually come back.
Mistake 5: Never saving what works
The biggest leak. You write a great prompt once, get a great result, and six weeks later you're reinventing it from memory. That's pure waste.
The fix — keep a personal prompt library. Every prompt that works twice gets saved with a one-line note on when to use it. After a few months you stop writing prompts from scratch entirely — you reach for a tested one and fill in the brackets.
This is, honestly, the difference between people who "use AI" and people who get 5+ hours a week back from it.
Putting it together
The throughline in all five fixes is the same idea: treat the prompt as a specification, not a question. Role, constraints, output shape, context, iteration, and reuse. None of it is advanced — it's just disciplined.
If you want to skip the trial-and-error, I put together a free PDF with 10 of my most-used prompts — email triage, code review, meeting summaries, decision matrices, content repurposing, and more. Each one has the placeholders marked and a note on why it works:
👉 10 AI Prompts That Save Me 5 Hours Every Week (free PDF)
And if you want to go deeper than copy-paste — the underlying patterns, few-shot and chain-of-thought techniques, domain playbooks for coding and writing — I built a full Prompt Engineering Masterclass and a library of 330+ prompts for exactly that.
But honestly? Start with the free one. Fix mistake #5 today — grab the pack, save the ones that fit your workflow, and stop reinventing them.
What's the one prompt you use every day that you wish you'd found sooner? I'm always adding to the collection.
Top comments (0)