Give the assistant a brief before asking for prose
State the reader, purpose, source material, length and tone. Tell it which facts it must preserve and what it must not invent. Ask for a structure first if the document is long; it is cheaper in effort to fix the outline than to rewrite an irrelevant finished draft.
For an email, define the action you want the recipient to take. For a summary, say whether the reader needs decisions, risks or a neutral account. For a translated draft, specify the language variety and preserve names, units, dates and quoted terms.
Draft a concise email for [reader] asking for [action]. Use only the facts below. Preserve amounts and dates exactly. If a required fact is missing, list a question after the draft instead of making it up. Tone: direct and considerate.
Separate the facts from the wording
Make a fact ledger before drafting: names, amounts, dates, commitments and source links. Keep it beside the generated text. The assistant may improve the wording while changing a condition, turning a possibility into a promise or assigning an action to the wrong person.
Do one pass for facts and a separate pass for style. Requesting a more confident tone should not change the strength of the evidence. Remove unsupported statistics, customer claims and quotations even when they make the paragraph sound convincing.
Worked example: preserve a missing deadline
Fictional source note: 'We discussed moving the workshop. Priya will check room availability. No new date was agreed.' A misleading summary would say: 'Priya will move the workshop to next Friday.' It invents a deadline and turns a check into a booking.
A faithful revision is: 'Priya will check room availability for a possible workshop move. The new date is not decided.' It preserves the action and uncertainty. This example illustrates editorial checking; it is not output attributed to a product or evidence that one model is better.
Compare this draft with the source notes. List every added fact, changed deadline, stronger promise or missing condition. Suggest a correction for each. Do not rewrite unrelated sentences.
Edit for a reader, not for an AI style score
Read the draft aloud. Replace vague benefits with the concrete action or example the reader needs. Cut repeated introductions and conclusions. Keep the terms your audience actually uses, and explain unavoidable specialist language on first use.
You can ask for three different openings or a shorter paragraph, then choose and revise yourself. A generated option is a suggestion, not an obligation. Preserve your own judgment about what belongs in the document.
- Purpose: does the opening tell the reader what matters?
- Facts: do names, numbers and conditions match the source?
- Voice: would you say this to the intended reader?
- Clarity: can any sentence be simpler without losing meaning?
- Action: is the next step explicit and actually authorized?
Common questions
- Can I send an AI-generated email as it is?
- Review the facts, tone, recipient and intended action first. A draft can invent a commitment or omit a condition even when the wording is polished.
- How do I keep my own writing voice?
- Give a clear brief, select only useful suggestions and edit the result for the actual reader. Keep your judgment over the final structure, examples and wording.
AI-assisted drafting and source review by Agent Search Engine. Fictional examples and proposed tests are labeled. Read our editorial method or report a correction.
More practical guides
Everyday AI / Decision guide
How to choose an AI tool for your everyday work
Choose AI tools by the job, setup, data, language and total cost. Compare two candidates with a free worksheet and practical tests, without coding.
Read the guideEveryday AI / Decision guide
How to research and learn with AI without losing the sources
A practical AI research workflow: define a question, collect sources, verify citations and separate evidence from inference. Includes a worked source-checking exercise.
Read the guideWorkflow / Decision guide
How to build AI workflow automation
Most “agents” should be workflows with LLM steps. Triggers, the engine, the LLM step, approval gates, reliability — and when to graduate to an agent loop.
Read the guide
