When the First Draft is Wrong: Exactly What to Type to Fix AI Output

You've been using AI long enough to know the first draft is rarely the one you send. The email comes back sounding like a compliance circular. The status update reads like it was written for a board presentation, not a Slack message to your reporting manager. The performance review draft has the word "consistently demonstrates" three times.

So you type something like: "Make it better." Or "Fix the tone." Or "This feels too generic." And you get almost the same thing back.

The problem isn't the AI. It's the feedback. Vague corrections produce vague improvements.

Why "Make It Better" Doesn't Work

When you tell an AI to "make it better," the model has no direction. It interprets "better" as "more" — more detail, more structure, more polish. "More professional" means "more formal." "More human" means "add contractions."

The fix: describe the specific problem AND the desired outcome. Not "make it shorter" but "cut this to 3 bullet points, each under 15 words." Not "fix the tone" but "rewrite this so it reads like a message to a colleague, not an official memo."

The Six Ways AI Drafts Go Wrong

Before fixing the draft, you need to name what's actually broken. Most bad AI output falls into one of these six categories.

1. Too Long and Padded

The model adds length to appear comprehensive — restating the same point in different words, adding "context" paragraphs that don't add information, ending with a summary of what was just said. A request for a quick status update returns four paragraphs. How to spot it: Read the first and last sentence. If they say the same thing, everything in between is padding.

2. Sounds Like AI

AI writing has recognizable patterns: em dashes used at 2–3× the rate of human writers; transition phrases like "Furthermore," "Moreover," "It's worth noting that"; grandiose words like "paradigm," "synergy," "ecosystem," "holistic," "tapestry"; the "It's not X — it's Y" construction; every sentence landing in the same length range with no variation; and filler openers like "Great question," "Certainly," "Absolutely."

3. Wrong Tone for the Context

AI defaults to a "safe professional" mode — the equivalent of business casual for writing. The problem: it applies the same tone to a message to your team's WhatsApp group and a letter to your company's CFO. It doesn't read the room.

4. Missing the Point

A documented failure mode called context drift: in long conversations, the model gradually loses track of the original intent. You ask for a status update on a delayed vendor delivery. The AI writes a general update about the project without mentioning the delay or what's blocking it.

5. Wrong Audience

AI writes for a generalist reader — the default mode is "explaining this to someone who needs background context." If you're writing a quick note to a reporting manager who already knows everything about the project, you don't need to re-explain the project.

6. Doesn't Sound Like You

Without samples of your writing, AI defaults to a template voice — polished, generic, interchangeable. Over time, colleagues who know your writing style notice something is off without being able to say exactly what.

The Exact Phrases That Work

For Length Problems

Cut this to 3 bullet points, each under 15 words.

Remove anything that restates what was already said. Keep only the new information.

This should be under 100 words. Cut the context paragraph — the reader already knows the background.

Shorten this to 5 sentences maximum. Remove the summary at the end.

For Tone Problems

Rewrite this in a more conversational tone. Sound like a colleague sharing news, not a company making an announcement. Use contractions.

Casual but professional. Short sentences. Like a message to someone I've worked with for two years, not a formal memo.

Make this 30% less formal. It's going to a peer, not a senior leader.

This reads like a press release. Rewrite it like a quick update between colleagues.

For AI-Sounding Patterns

Do not open with any acknowledgment phrase or filler word.

Do not close with a summary or a restated thesis.

Remove these phrases: furthermore, moreover, it's worth noting, this highlights, as mentioned above.

Vary the sentence length. Mix 5-word sentences with 25-word sentences in the same paragraph.

Remove any word that ends in "-ize" or "-tion" if a simpler word exists.

For Wrong Audience

The reader is my reporting manager who already knows the full project background. Remove all context-setting. Lead with the status and what's needed from them.

This is for a VP who reads 200 emails a day. Lead with the decision or action needed. Everything else is secondary.

The reader is a senior engineer, not a customer. Remove the hand-holding explanations.

For Missing-the-Point Problems

The actual point of this email is: [one sentence]. Make sure that sentence appears in the first two lines.

What I actually need the reader to do after reading this: [specific action]. Make that clearer.

You've written about the project generally. I need this to specifically address: [the delay / the blocker / the decision needed].

The Contrast Method: Show, Don't Tell

Instead of describing what you want, show the AI an example. This is especially effective when you struggle to articulate what's wrong.

How to use it: Paste the AI draft, paste 2–3 sentences of your own writing (from a past email, a message you sent recently), then ask: "Rewrite the draft to match the tone and sentence structure of my writing sample below."

Research on this technique shows 5–10% accuracy improvement on complex writing tasks compared to instruction-only corrections. The model picks up rhythm, vocabulary choices, and structural patterns from examples in ways that descriptions alone don't capture.

A simpler version when you don't have samples: paste one sentence that has the right feel and ask "Write the rest in this style."

The Self-Critique Loop

What are the 3 biggest weaknesses in this draft? For each one, name the specific problem and what it should be instead. Then rewrite the draft fixing those three things.

This works because it forces the model to reason about the output before revising. Important caveat: Self-critique can backfire on already-solid drafts — the model may invent problems that don't exist. Use this when you know something is off but can't articulate exactly what.

Workplace-Specific Fixes

Status Updates to Reporting Managers

What goes wrong: AI inflates language and buries the status. "Made significant progress on the Q2 initiative" instead of "Finished sections 1–3, blocked on vendor input for section 4, need your call on the timeline by Thursday."

This status update is for my reporting manager. Rewrite it in this format only:
Done: [what was completed]
In Progress: [what I'm working on now]
Blocked: [what's stuck and what I need from them]

Maximum 3 bullet points per section. No filler phrases. Be specific about blockers and what action is needed.

Emails to Senior Leaders

What goes wrong: AI writes for a blog audience — assumes the reader needs context, buries the main point in paragraph 3, uses inspirational language that feels performative.

The reader is a VP who gets 300 emails a day. Lead with the specific decision or action needed in the first sentence. Context goes second, only if essential. Remove anything that isn't needed for them to act. Maximum 5 sentences total.

Performance Reviews

What goes wrong: Generic praise appears everywhere — "consistently demonstrates excellence," "goes above and beyond." Every review starts sounding the same.

Rewrite this performance review using only the specific facts I give you. Do not add generic praise. Do not use the phrases "consistently demonstrates," "goes above and beyond," or "team player."

Here are the specific facts:
- Accomplishment 1: [specific project, specific outcome, specific number if possible]
- Accomplishment 2: [same format]
- Area for development: [specific, not vague]

If there are gaps in the information, flag them rather than filling them with vague positives.

Meeting Summaries

What goes wrong: AI writes a narrative recap of what was discussed — not what was decided. Captures the conversation, misses the outputs.

Rewrite this meeting summary in this format only:
Decisions Made: [list only things that were actually decided]
Action Items: [person name — what they're doing — by when]
Open Questions: [things that came up but weren't resolved]

No narrative prose. Do not summarize discussion that didn't produce a decision or an action item.

The "Write Like Me" Technique

If you use AI regularly for writing, the most time-saving setup is teaching it your style once — so you spend less time correcting every individual draft.

Step 1: Paste 300–500 words of your own writing into the chat. Use emails or documents you wrote when not using AI. Step 2: Ask: "Describe my writing style in bullet points. Focus on: sentence length, vocabulary level, use of contractions, formality, structure, and any recurring patterns." Step 3: Read the description. Correct anything that's wrong. Save the accurate version. Step 4: For future drafts, paste this style description at the start of your prompt: "Write in this style: [your style description]."

ChatGPT's Custom Instructions feature (Settings → Personalization) lets you save this permanently. Claude Projects can store a style guide you reference across multiple conversations.

If you want to understand why some AI tools produce better-quality first drafts and require fewer correction rounds on writing tasks, this comparison of Claude vs ChatGPT for writing covers the specific quality gaps that show up in professional writing.

The Real World Story

You've been here. You got a bad draft, typed something like "make it better" or "fix the tone," and got almost the same thing back with slightly reshuffled sentences. Maybe you tried a few more rounds and gave up, deciding that editing the AI's output is more work than just writing the thing yourself.

That conclusion is wrong — but the feedback method was the problem, not the tool. The first time I applied the "Correct with a hard constraint" approach on a performance review draft — "Rewrite this using only the specific facts I give you. Do not add generic praise." — the output went from something I would have thrown away to something I actually sent. One correction round, thirty seconds. The difference was naming the specific problem and setting a bounded direction.

The mistake most people make is treating AI correction like asking a person to "try again." People can read between the lines; AI can't. Every correction needs to tell the AI exactly what went wrong and exactly what direction to go instead. If you can answer those two questions, most drafts are fixable.