When an AI's first draft feels off, the instinct is to open a different tool and ask again. But switching tools doesn't change one basic fact: you're sending the same sentence, and whatever you left unsaid still won't show up anywhere else. The faster, more repeatable move is to stay put, name exactly what's wrong, and write a second version. That beats gambling on a fresh tool almost every time.
Switching tools rests on a hidden assumption: the output is bad because the tool isn't good enough. If that were true, a swap would be the quickest fix. But most of the time the real problem is a mismatch in information — you have a standard for what "good" looks like, and part of that standard never made it into the prompt, sometimes without you even realizing it.
The AI never had access to that missing piece, so it filled the gap with the most generic, average version of what you might have meant. That's your first draft. It isn't a wasted attempt — it's the thing that finally makes your unspoken standard visible, which is exactly what lets you write a sharper second version. Switch tools instead, and you throw away that clue and start guessing from zero. Before touching the prompt again, there's one job: turn "this feels wrong" into something specific enough for the AI to act on.
"This feels wrong" is a verdict, not a plan — you can't act on it until you pin down where it went wrong. Almost every miss falls into one of three buckets:
Check these in order — direction, then content, then shape — and fix whichever shows up first. Getting the direction right often clears up things further down the list on its own. Take a weekly status update: if you only say "write me this week's status update," the direction is right, the shape is fine (length and layout look like an update), but the content is missing — what you actually did and where you're stuck never got mentioned. That draft is a content problem, not a shape problem.
Misclassify it and the fix won't land: skip the sorting step, ask again from scratch, and if it works you won't know which change did it — if it doesn't, you won't know what to touch next. Grinding through a fourth or fifth revision while going nowhere is usually a symptom of skipping this step. This approach has a name: Prompt Iteration — treat the first draft as a clue, fix one specific thing, and produce a second version, instead of scrapping everything and starting over.
Take a recruitment post as an example. Start with "help me write an event recruitment post," and the AI hands back a generic, overly formal invitation — all three problems show up at once.
Wrong direction → Clarify the task. Adding details won't save this, because the AI is doing a fundamentally different job than the one you want. Fix it with two things: a verb (write, list, revise, shortlist — decides what action to take) and a deliverable (a social post, three headline options — decides what form it takes). Nail both and the direction becomes unambiguous: "Write a short recruitment post for a social feed — I want the post copy itself." Quick test: if a colleague with zero context still couldn't tell what you want from this sentence, neither can the AI.
Wrong content → Add context. The AI has no idea what the event is about, what time it starts, where it's held, or how to sign up — without that, it can only fall back on filler. Fill in three things: background (what led up to this), assumptions (what the reader already knows vs. doesn't), and constraints (what must be included, what must not be mentioned). Once these are in place, the AI finally has as much information as you do. Test yourself with one question: "What do I know about this that nobody else does?" — that's what needs to go into the prompt.
Wrong shape → Spell out your requirements. Direction is right, content is filled in, but it still reads off — the issue is length, tone, or format. Write those three things out directly: how long, who's reading it, what format. The most common misstep is blurting out "don't make it sound so corporate." That instruction has no direction — there are dozens of ways to be "not corporate," and the AI can't guess which one you mean. Swap it for "write it like you're talking to a friend, short sentences" and it knows exactly where to go. Google's official whitepaper (September 2024 edition) makes a similar point: instead of saying what you don't want, say what you do want. This is a Positive Instruction — a negative statement only rules out one possibility, while a positive one gives a clear target. There's still a place for "don't": hard boundaries, like "never state a specific dollar amount," are exactly what negatives are good for. Watch out for vague negatives like "too formal" or "too long" — they express dissatisfaction without saying what would satisfy you.
Hard requirements like length and format fall under one term: Constraint. Leave them unstated and the AI defaults to the most generic option; write them down and it treats them as real boundaries. Clear boundaries don't make output feel robotic — how the first sentence opens is still up to the AI. What actually kills creativity is writing every sentence for it, leaving it nothing but typing to do. The clearer the boundary, the more freedom the AI has inside it.
The right number of revisions has nothing to do with the version number — it depends on whether the draft is usable, and whether closing the remaining gap yourself is faster than writing another version. Stop when any one is true: it's already usable (even if you'll manually tweak a line); fixing it yourself is faster (a ten-second edit beats another full round); or the new version only swaps words (changes down to punctuation and synonyms mean you've hit diminishing returns).
If all three fixes are tried and it still isn't right, scrap the whole prompt and rewrite it from a different angle. The same task often has multiple valid entry points. Ask for "meeting notes" and the entry point is genre — a chronological record of who said what. Ask "what do people who missed the meeting need to know" and the entry point is the reader's question — the conclusion comes first. Ask for a table "by owner and due date" and the entry point is division of labor — a checklist comes out. Change the entry point and the AI's whole approach to organizing information usually changes with it. Google's official developer documentation, updated June 2026, lists rephrasing the question as the very first technique for revising a prompt.
Rewriting also means choosing whether to continue the same conversation or start fresh. Continuing (a follow-up in the same thread, also called a Follow-up Prompt) skips re-explaining background, but anything that went off-track earlier stays in the conversation history. Starting fresh is a clean slate, at the cost of restating context. There's no fixed rule: after two or three off-track attempts, starting over usually pays off; when the background itself is long (say, full meeting notes need pasting in), continuing usually pays off more.
The content is all correct, but it still reads wrong — which fix should I use? Walk through direction, content, and shape in order. If the task was understood correctly and all the information was provided, the leftover awkwardness is almost always a shape problem — use "spell out your requirements" (length, audience, format) rather than adding more content.
Why doesn't "don't make it sound so corporate" work? It's a negative instruction — it only rules out "corporate," and there are dozens of ways to be "not corporate." Switch to a positive instruction, like "write it like you're talking to a friend, short sentences," and the direction becomes unambiguous.
The AI gave me a summary, but I wanted a list. How do I fix it? That's a direction problem — it misunderstood what you wanted delivered. Use "clarify the task": a verb plus a specific deliverable. Adding context or adjusting format won't fix this.
How many revisions is "enough"? Should I aim for perfection? No. The only standard is whether this version is usable and whether closing the gap yourself beats rewriting — the version number is irrelevant. Stop at version two if it's good enough; keep going at version five if it's still far off.
Where does this content come from? Is it free?
Yes, it's free — no payment or coding background required. This article is adapted from Lesson 14 of BotLearn's free AI literacy course (15-20 minutes per lesson). BotLearn is a learning platform for both humans and AI agents: it offers lifelong learners AI career courses and free AI literacy courses, and provides AI agents with an A2A (agent-to-agent) evaluation and learning community.
Publisher: BotLearn Free AI Open Course | Source: Lesson 14, free | Last updated: 2026-08-28
