Iterating with AI: How to Steer Revisions Accurately

When a draft misses the mark, many users reflexively reply: "Please optimize this." In practice, it is a gamble. You might be looking to add a sign-up link, but the AI hears "optimize" and rewrites the phrasing, rearranges the sentences, or alters the tone.

You can only evaluate whether a revision succeeded if you clearly specify what to change and what to keep.

When facing several issues, focus on one primary goal and clearly define three things:

• Preserved elements: Which parts are already correct and must not be changed;
• Specific revision action: What to add, delete, or modify in this turn;
• Acceptance criteria: What the updated version must achieve to pass review.

Each of these takes a sentence or two: pick one focus for this round, lock the parts that are already right so the AI doesn't quietly rewrite a correct date or place, turn the complaint into a concrete action, and say what counts as done.

When the second draft comes back, compare it against all three before you approve it, since a cleaner tone only shows the style improved.

01 Priority: Focus the Revision Target

When faced with multiple flaws (for methods to spot gaps in a first draft, see Reviewing AI Drafts: Catching Omissions and Fluff), ask yourself: which issue most prevents this content from being usable right now? That gives you an objective priority.

In our community reading announcement, the first draft said: "Saturday at 3:00 PM in the activity room for a Family Reading Session." It omitted the end time, audience age, preparation, and sign-up steps, while including fluffy pleasantries. While the fluff takes up space, the missing logistical details leave parents unable to register. Factual completeness naturally comes first.

When prioritizing edits, this general hierarchy serves as a helpful baseline:

1. Factual errors and ungrounded additions, which mislead readers;
2. Critical omissions, which leave readers unable to understand or act;
3. Poor formatting or sprawling paragraphs, which make information hard to find;
4. Minor tone and phrasing adjustments.

Priorities always depend on how the output will be used. If you are drafting a sensitive apology note, an abrasive tone will cause damage even if all the facts are technically correct. In that context, tone adjustments take precedence.

Once priorities are set, frame your revision around a single clear goal: adding the necessary logistical details. Anthropic's guide on chaining complex prompts notes that packing unrelated goals into one turn often causes models to drop constraints. Combining multiple related tweaks under one coherent goal is fine, provided you do not introduce conflicting requirements.

02 Preserved Items: Lock Down Correct Content

One of the most frustrating aspects of working with AI is asking it to add a single missing sentence, only to have it rewrite the entire text and discard previously accurate details. AI does not have an instinct to "leave well enough alone." Before requesting changes, identify what the draft already got right and state that those parts should stay untouched.

In our draft, separate what is accurate from what needs improvement:

The same principle applies to tone and formatting. If the draft's length, tone, or layout was already appropriate, state that explicitly so the model preserves those qualities during the revision.

03 Actionable Fixes: Turn Complaints into Steps

Complaining simply points out what is wrong; a revision instruction tells the AI what to do next. Throwing a vague critique at the model:

Problem: The announcement is incomplete.

forces the AI to guess what is missing. Refer to your source notes, list the specific items to add, and make the next step unambiguous:

Add the end time (4:00 PM), target audience (children aged 6 to 10 and parents), preparation (bring a favorite book; arrive 10 minutes early), and registration (sign up in the group chat by Friday, 6:00 PM).

The clearer the action, the easier it is to review the result. If the previous version only had 3:00 PM, you check whether the new version writes 3:00 PM to 4:00 PM. Apply this pattern whether you are correcting facts, cutting filler, or restructuring paragraphs: name the target and state the change.

Locking in accurate facts before supplying missing information prevents the AI from fixing one detail while breaking another.

04 Boundaries: Define What Passes Review

Naming the revision action is half the equation; you also need to define what qualifies as a successful result. Google's Gemini prompt design strategies describe this as refining prompts through iterative constraints: identify gaps in the output and tighten boundaries around length, structure, and required elements. In our example, we want complete details without letting the notice turn into an essay:

• Time range, audience, preparation, and sign-up method are all present;
• Event title, start time, and location remain intact;
• Output remains a single paragraph under 120 words.

Combining preserved elements, revision actions, and acceptance criteria produces a focused follow-up prompt:

Keep the event name and start time. Update the location to "Community Activity Room."
Add the end time (4:00 PM), target audience (children aged 6 to 10 and parents), preparation (bring a favorite book; arrive 10 minutes early), and registration (sign up in the group chat by Friday, 6:00 PM).
Write as a single paragraph under 120 words.

When the new draft arrives, check it against these three instructions, using your original notes as the source of truth. (If you start a fresh chat session for revisions, remember to provide the source material, previous draft, and revision requirements together so the model has the full context.)

05 Comparison: Check Versions on Equal Ground

Using the prompt above, the AI produces this updated draft:

This Saturday from 3:00 PM to 4:00 PM, a Family Reading Session will be held in the Community Activity Room for children aged 6 to 10 and their parents. Please arrive 10 minutes early with a favorite book. Please sign up in the group chat by Friday, 6:00 PM.

Rather than approving it simply because the writing looks cleaner, compare it against your criteria:

Comparing the two versions confirms that omissions were filled, accurate details were kept, and the length fits the channel. The follow-up achieved its purpose. Finally, check whether compressing the text created any secondary issues.

If a detail is still missing after a round of editing, refine that specific requirement. If a new issue appears, address it in the next turn. After each iteration, check that earlier requirements haven't slipped away. If you find yourself needing facts you don't have, pause and confirm them before re-prompting.

Comparing iterations shows whether a revision worked. Deciding whether the text is safe to distribute requires a final review against real-world risks; see Before Using AI Output: A Four-Step Final Checklist.