Lesson 004Field guide · Prompting 101
How to Improve Bad Outputs
Iterate instead of restarting: turn a weak answer into a strong one.
You'll learn
- Why the first answer is a draft, not a verdict
- The four diagnoses: too generic, wrong tone, wrong format, wrong assumptions
- The follow-up move that matches each diagnosis
- How to keep what works while fixing what doesn't
- When to repair the conversation — and when to restart it
01Section
The first answer is a draft
When an AI answer disappoints, most people do one of three things: retype the same prompt and hope, switch tools, or give up. All three throw away the most useful thing you now have — a concrete example of what you didn't want.
The better move is to treat the first output as draft one and reply to it. AI tools are conversational: they take direction mid-stream, and a targeted follow-up is usually faster than a better first prompt would have been to write.
A useful analogy
It's like getting a first draft back from a fast but literal-minded assistant. You don't fire them and hire a new one — you hand the draft back with a note: "good structure, wrong tone, and you assumed our customers are companies; they're families."
Key takeaway
You don't need a perfect prompt. You need a decent prompt plus one good note.
Checkpoint
02Section
Diagnose before you fix
"Make it better" is the follow-up equivalent of "write something about marketing" — a vague complaint gets a vague revision. Before you type anything, name what's actually wrong. Four diagnoses cover almost every weak output:
Too generic
Reads like it was written for anyone. Polished, plausible, and says nothing specific to you.
Wrong tone
The content is fine but the voice is off — too stiff, too cheerful, too salesy.
Wrong format
Good ideas in the wrong shape: an essay when you needed a list, ten options when you needed one.
Wrong assumptions
It guessed something about your situation — audience, industry, goal — and guessed wrong.
Checkpoint
03Section
Match the repair to the diagnosis
Each diagnosis has a matching repair. The pattern is always the same: name what's wrong, then supply what was missing.
Too generic → add the specifics: "Rewrite this for my actual situation: a 12-person accounting firm whose clients hate jargon." Give it the details it was forced to guess.
Wrong tone → show it, don't describe it: "Less press release, more text message." Or paste a sample: "Match the tone of this: ..." A reference beats another adjective.
Wrong format → restate the shape: "Same content, but as a table with columns for cost and effort." The ideas survive; only the container changes.
Wrong assumptions → correct the record: "You assumed my customers are businesses — they're families with young kids. Revise with that in mind."
The weak output you got
"Boost your productivity with these universal tips: wake up early, make a to-do list, minimize distractions, take regular breaks..."
The repair you send
Too generic — my readers are freelance designers juggling several client projects at once. Rewrite with tips specific to client-driven schedules, and cut anything that would apply to literally anyone.
Diagnosis named, missing specifics supplied, one clear instruction. That's a repair — and it usually lands in one pass.
Checkpoint
04Section
Say what to keep
A follow-up that only criticizes has a side effect: the AI often rewrites everything, including the parts you liked. The fix is to fence off the good parts explicitly — every strong repair says what stays as well as what changes.
A repair that protects the good parts
Keep the list format and the friendly tone — both are right. Replace ideas 2 and 4, which are too expensive for a small shop, with two cheaper alternatives. Leave the rest exactly as it is.
Checkpoint
05Section
Repair or restart?
Repairs work when the foundation is right and one dimension is off. But sometimes the first prompt aimed at the wrong target entirely — you asked for a marketing plan when you needed help deciding whether to market at all. No follow-up fixes a wrong destination.
Long repair threads have a second problem: after several rounds of contradicting instructions, the AI starts blending old directions with new ones, and each revision gets muddier instead of sharper.
A rule of thumb
Two focused repairs. If the third message would be another repair, open a fresh chat and write a better first prompt instead — you now know exactly what it needs to say.
That's the part beginners miss: a restart isn't starting over. Everything the failed thread taught you — the missing context, the tone reference, the format, the things to avoid — goes into the new opening prompt. This loop, run a few times, is how your prompts actually get good.
Pause and think: Think of the last time an AI answer frustrated you into giving up. With the four diagnoses in hand, what would your one-sentence repair have been?
Checkpoint
Prompt exercise
Run it, grade it, repair it
Copy this prompt into ChatGPT, Claude, Gemini, Copilot, or whichever AI tool you have access to — the website doesn't run AI itself. The first answer will probably be decent but flawed, and that's the point. Diagnose it with the four labels (too generic, wrong tone, wrong format, wrong assumptions), then send one repair that names the diagnosis, says what to keep, and supplies what was missing. Compare the revision to the original.
Write a short welcome message for new subscribers to my email newsletter. The newsletter shares one practical tip each week. Keep it under 120 words.
Reflection: Which diagnosis fit your first output? And how much did one targeted repair change — compared with what retyping the same prompt would have gotten you?
Quick check
4 quick questions — no pressure
There's no pass or fail here. Answer them all, and we'll show you the answers either way.