TL;DR
- AI is good at two very different jobs: finding things in what you've already written, and inventing things that sound plausible. Only the first belongs in your career record.
- A resume line that's technically written well but generated, rounded up, or "made to sound better" fails the moment someone asks a specific follow-up question.
- Retrieval means searching your chats, docs, and notes and surfacing what's actually there — including the parts that are messy or incomplete.
- Invention means guessing at a number, smoothing over a gap, or writing a sentence that merely sounds like something you'd have done.
- A simple test before accepting anything AI drafts about your work: could you defend this sentence, unprepared, to the person who was actually there?
- The best prompts for this ask explicitly for retrieval and explicitly forbid invention — "don't infer or guess, flag anything you're not sure about."
Two different jobs wearing the same interface
Retrieval and invention look identical in a chat window — the model produces text either way, in the same confident tone. What actually separates them is the source: did the model pull this from something you actually wrote, or generate something new that merely resembles it?
Where invention creeps in
"Make my resume sound better" is an open invitation to invent — round up a metric, add impact language that isn't grounded in anything specific, exaggerate scope just slightly. It reads fine right up until someone in an interview asks how you measured it, and there's nothing real behind the sentence.
What retrieval actually looks like
Point the AI at your real chat history, tickets, documents, and reviews, and ask it to extract — with a source — not to summarize a wish. A tool doing this well should flag uncertainty instead of filling gaps with a plausible-sounding guess, and should be able to point back to where a claim came from.
The defend-it-live test
Before you accept anything AI drafts about your own work, ask one question: could you explain this claim, unprompted, to the person who was actually in the room? If the honest answer is "not really," it's invention — however well it reads.
Copy this prompt
- “Go through everything I've shared with you and pull out real moments of my work — wins, decisions, challenges, feedback. For each one, tell me what happened, what I did, and what resulted, based only on what's actually in what I've shared.”
- “Don't infer, guess, or round anything up. If something is unclear or you're not sure it counts, flag it as uncertain instead of stating it as fact.”
- “I'll verify and rewrite everything myself afterward — your job here is finding, not writing.”
Frequently asked questions
Isn't it fine to have AI polish the wording, even if it invents a little?
Polishing wording is fine. Inventing the substance underneath it — a number, a scope, a claim of ownership — is what causes trouble later, usually at the worst possible moment: a follow-up question in an interview or a review.
How do I know if something an AI drafted is retrieval or invention?
Ask where it came from. If it can point to a specific message, ticket, or document, it's retrieval. If it can't, or it hedges, it's invented — even if it sounds confident.
What's a good prompt to keep AI on the retrieval side?
Something like: “Only include things you can point to in what I've shared. If something's unclear or you're not sure it counts, flag it as uncertain instead of stating it as fact.”
Does this mean I should never let AI write anything from scratch?
No — drafting phrasing, structure, or a first pass at a sentence is fine. The line is at facts: what happened, what you did, and what the result was should always trace back to something real.
Related reading: How to build a brag doc with AI.