Generic AI writing is usually a symptom of weak inputs and weak review. The prompt names a topic but not a reader decision. The draft receives no source packet, so it fills gaps with broad claims. Voice is described with adjectives instead of examples. The final instruction asks for polish without defining what must remain true.
A surface rewrite can vary the phrasing. It cannot supply missing evidence or decide which claim should survive. The repair therefore begins before sentence-level editing.
1. Diagnose the missing control
Read the draft once without editing. Mark four things:
- Claims: facts, numbers, dates, comparisons, promises, quotations, and causal statements.
- Abstract benefits: phrases such as improve efficiency or unlock potential that do not name a visible change.
- Reader gaps: sections that never state who is deciding what.
- Voice drift: sentences that could appear unchanged on an unrelated brand's page.
This diagnosis prevents a common failure: making every sentence more varied while leaving the draft equally unsupported.
2. Build a source packet
List the facts the draft may use, their links, the date checked, proof examples, missing evidence, and claims that are not allowed. Keep prices, policies, performance numbers, and quotations in separate rows because they can age or lose context quickly.
ChatGPT's own help materials warn that generated answers can be incorrect and should be checked when accuracy matters. Treat model output as a draft, not as the source trail. See OpenAI's accuracy guidance.
3. Restore the reader's job
Replace the topical opening with the specific decision or constraint. A weak opening says that AI writing is changing content creation. A stronger opening says that the editor has twenty minutes to verify a draft before it reaches a client.
The stronger version is not merely more vivid. It creates an acceptance test for every later paragraph: does this help that editor make the draft safer or faster?
4. Repair the strongest claim first
Find the sentence that carries the most commercial or factual weight. Support it with a source, a worked example, or a narrower claim. Do not spread limited proof across several ambitious statements.
Before: This workflow dramatically improves conversions. After: This workflow puts the source, proof example, and checkout action in one review sequence. The current test has not established a conversion lift.
The repaired version gives the reader something observable and preserves the uncertainty that remains.
5. Use voice evidence, not costume
Provide two to four representative passages and extract a few visible rules: sentence length, level of directness, preferred concrete language, and patterns the brand refuses. Ask the model to explain the important changes so a human can review whether the rules were applied for the right reason.
A long list of words to ban can make prose stranger without making it more recognizable. Positive examples and explicit reader context are more useful than blanket substitution.
6. Run the final six checks
- Every important claim has support.
- The reader and decision are visible near the opening.
- Abstract benefits have concrete conditions or examples.
- Voice matches the supplied samples.
- Observed evidence is separated from hypothesis or promise.
- One useful next action follows naturally from the proof.
Use the free cleanup checklist on a live draft. For the full repair chain, inspect the public before-and-after examples before considering the paid kit.
Why the Bureau archives this as a method
The reusable mechanism is not a particular prompt. It is proof-led content: source control, visible repair, a useful free artifact, an owned-audience path, and a decision rule. That mechanism can survive model changes and product formats, so it belongs in method record VB-M002.