Facts and Prose: Which Resume Fields AI Can Touch

A resume contains two different kinds of content, and treating them the same is what causes most AI-assisted errors. Fixed fields — employer names, dates, titles, qualifications, credentials, numbers — either match an external record or they do not, and no tool should ever generate or reformat them. Prose — bullet phrasing, summary, ordering, emphasis — is where a language model is genuinely useful. Draw the line explicitly and most of the risk disappears.

The fixed fields

These should be typed by you, from your own documents, and read separately on the final pass:

Employer legal names, including agency names where you were placed.
Start and end dates, to the month, from payslips or letters.
Job titles as recorded by the employer.
Qualification names, awarding institutions, and completion status.
Certification names, issuing bodies, and expiry or lapse status.
Any number: team sizes, budgets, percentages, user counts, years of experience.
Contact details.

What these have in common is that a third party holds an authoritative version. Editing them for flow is meaningless, because there is nothing to improve — there is only agreement or disagreement. And of the two, dates are the field most reliably checked against someone else’s records.

The prose

These are the parts where wording is a real choice and a tool can help:

Bullet phrasing, once the underlying claim is fixed.
Cutting length without losing the substance.
Ordering bullets and sections for a particular application.
The summary, written from the finished page rather than from a general description of you.
Consistency of tense and structure across entries.
Spotting vagueness — asking a model which of your bullets could describe anyone is a genuinely good use of it.

In every case the tool is operating on material you supplied, not supplying material. That is the whole distinction.

The workflow the split implies

Write the fixed fields yourself, from documents. Employers, dates, titles, credentials.
Write each bullet’s claim in plain language, from memory, badly. One sentence, no polish. This is the step people skip and it is the one that keeps the claims yours.
Then use a tool on the plain sentences: tighten, shorten, remove repetition, suggest a clearer order.
Check the output against your plain versions, looking specifically for added specifics, escalated verbs and widened scope.
Re-enter the fixed fields last, or verify each one against your source, because reformatting steps quietly normalise them.
Write the summary last of all, from the finished page.

Step five is not paranoia. Ask a tool to reformat an entire resume and it will treat titles and dates as text like any other, standardising a grade into a conventional title or dropping a month. The change is invisible because those lines look like structure rather than claims.

What to ask a tool, and what not to

Useful requests:

“Shorten this bullet without removing the specifics.”
“Which of these bullets could describe almost anyone?”
“Is this summary supported by the experience below it?”
“Rewrite this in plainer words.”
“What questions would you ask about this bullet?”

Requests that produce fabrication:

“Make this sound more impressive.”
“Add metrics to these bullets.”
“Write a resume for a senior data role based on this profile.”
“Fill in the gaps in my experience section.”

The second group all ask for content the model does not have, and it will supply it anyway. The distinction is not about phrasing the prompt more carefully; it is about whether the request can be satisfied from the information you provided.

The one exception in the prose column

Verbs sit awkwardly across the line. They are prose in form and factual in content, because the verb you choose is a claim about your role that another person could contradict. Treat the verb as a fixed field even though it looks like wording: choose it yourself, and if a tool changes it during a rewrite, change it back.

The same goes for scope nouns. “The onboarding flow” and “the customer experience” are not stylistic variants of each other.

Why this beats a general instruction to be careful

Telling yourself to check AI output carefully does not work, because the output is fluent and the errors are plausible. A structural rule does work: this column of the document is never generated, that column can be. It gives you something mechanical to verify rather than a judgement to make on every line.

It also makes the checking pass finite. You know which fields to read against source documents, and for everything else you are looking for the specific directions in which generated prose drifts rather than reading with undirected suspicion.

The residual work

The split does not remove the need to check the prose, because a well-phrased bullet can still overstate what you did. It removes the category of error that is hardest to spot and most damaging when found — a wrong date, a title you did not hold, a figure nobody supplied. Those are the ones that turn into a correction you have to send after the fact, and they are entirely preventable by refusing to let a tool near them.