Drafting, Editing, and Fact-Checking with AI
The concept
AI writing assistance works best when you treat it as three distinct modes, used in sequence, rather than one generic "write this for me" request. Conflating the three is where most disappointing AI-assisted writing goes wrong — asking for a finished, polished, fact-checked piece in a single prompt usually produces something that sounds finished but hasn't actually been checked.
Mode one: drafting from an outline. This is the fastest and roughest mode — turning bullet points, rough notes, or a structure into full prose for the first time. The goal here is speed and coverage, not final quality; a draft's job is to exist so there's something to react to and improve, not to be publication-ready. Good draft prompts hand the AI a clear outline or set of notes and ask for a complete first pass, explicitly framed as a draft ("turn these notes into a first-draft paragraph for each point — don't polish, just get the content down").
Mode two: editing an existing draft. This is a narrower, more controlled task: tightening tone, cutting length, improving clarity, or matching a specific voice — applied to text that already exists, rather than generating new content. Editing prompts work best when they're specific about what to change and what to preserve: "make this more confident, cut it by 30%, keep all the numbers exactly as written" gives the AI a clear target, whereas "make this better" invites it to rewrite more than you wanted, potentially changing facts or claims along the way.
Mode three: fact-checking the finished piece. This is the stage people skip most often, and it's the most important one. Ask the AI to list every specific factual claim, number, date, or quote in the piece — not to verify them (it can't reliably do that against its own output), but to surface them so you can check each one against a real source. AI-polished prose is a particular risk here: good editing makes writing sound more confident and authoritative, which can make an unverified or wrong claim feel more trustworthy than it actually is, not less. A clean, confident sentence is not evidence that the fact inside it is correct.
Running all three modes in sequence — draft, then edit, then fact-check — catches different problems at each stage: drafting catches missing content, editing catches tone and length problems, and fact-checking catches claims that need a source before anything goes out the door.
It's worth naming a fourth, smaller habit that compounds across all three modes: keep your original notes or outline next to the final draft, not just the AI's output. When editing tightens a sentence or fact-checking flags a claim, you want to be able to glance back at what you originally meant to say, rather than relying on the AI's increasingly-edited version as the source of truth for your own intent. This is especially important for anything with numbers, dates, or commitments in it — the kind of content where a small, confident-sounding drift during editing can turn into a real problem once it reaches a client or a colleague.
Worked example
A learner has rough bullet notes for a client update: "delayed two weeks, vendor issue, new date March 15, we absorbed the extra cost, client not happy but stayed." Draft stage: the AI turns this into a full first-draft paragraph covering all five points in plain sentences — rough, usable, not yet tuned for tone.
Edit stage: the learner prompts, "make this more confident, cut it by 30%, keep the March 15 date and the vendor-issue reason exactly as written." The AI tightens the paragraph into three crisp sentences, preserving the specific facts the learner flagged as untouchable.
Fact-check stage: the learner asks, "list every factual claim, date, and number in this paragraph." The AI returns: "(1) two-week delay, (2) vendor issue as the cause, (3) new date of March 15, (4) company absorbed the extra cost." The learner checks their own records and confirms all four — except the new date, which their calendar shows as March 17, not March 15 (the AI had subtly shifted the date during editing). They correct it before sending. Skipping the fact-check stage would have sent a wrong date to a client who was already unhappy.
Having kept the original notes alongside the final draft, the learner also notices the edited version dropped the phrase "client not happy but stayed" entirely during the 30% cut — a true and relevant fact that the AI judged as less essential than the others. The learner adds a short clause restoring it, since acknowledging the client's frustration directly was part of the point of the update.
Exercise
Take a real piece of writing you need this week — an email, a social post, or a section of a report. Run it through all three modes in order: draft from your own rough notes or outline, edit for a specific tone and length target you state explicitly, then ask the AI to list every factual claim, number, and date in the final version.
Manually verify at least two of the flagged claims against a real source (your own records, a document, a calendar) before considering the piece done. Compare the final draft against your original notes line by line, and note whether editing introduced any drift from your original facts, the way it did in the worked example above, or silently dropped a point you intended to keep — both are common and easy-to-miss failure modes.
Recap
- Treat AI writing as a three-stage pipeline — draft, edit, fact-check — each stage catches a different kind of problem.
- Be specific in edit prompts about what to change and what to preserve exactly, since editing can silently drift facts or drop content even while improving tone.
- Never skip the fact-check pass: confident, well-edited prose is not evidence the underlying claims are correct — always verify flagged claims against a real source, and check the final draft against your original notes.
Resources
- OpenAI Help Center: "Best practices for editing and fact-checking AI writing" — https://help.openai.com/
- Grammarly: AI writing and editing assistant — https://www.grammarly.com/ai
- Poynter: "A guide to fact-checking" — https://www.poynter.org/fact-checking/