How we found these
Resunova uses language models to analyse resumes and propose rewrites. Early on we shipped that output more or less as the model produced it, and the complaints that came back were not about tone or style. They were about the tool saying things that were not true: flagging a bullet for lacking numbers when it had four, presenting a rewrite identical to the original, listing a qualification as missing directly above an explanation of where the candidate met it.
So we built a validation layer that sits between the model and the user. Each check started as a specific observed failure. What follows is that list, because the same mistakes will show up in your own chat window, and none of them are hard to catch once you know the shape.
The general rule
The rewrite that changes nothing
The most common failure is not a bad rewrite. It is a rewrite that is not a rewrite at all, delivered with the same confidence as a real improvement.
A semicolon became 'and'. Presented as an improved bullet, with an explanation of why it is stronger.
This matters more than it looks. If you are working through fifteen suggestions and four of them are cosmetic, you spend your attention on noise and lose confidence in the eleven that are real. Our filter drops any rewrite that is byte-identical to the original after normalising whitespace, and any rewrite that overlaps the original above a word-similarity threshold.
Catch it yourself: ask what got better in one sentence. If you cannot answer, keep your version.
Quantified, with no number attached
Models know that resume advice says "add metrics". So they tag bullets as needing quantification, and describe their rewrites as more quantified, without ever adding a digit.
Labelled as adding quantification. 'Significantly' is not a quantity. The bullet contains exactly as many numbers as before, which is none.
There is a second-order problem here too: the rewrite also asserts an outcome ("improving ramp-up time") that the original never claimed. That is invention dressed as editing.
Our rule is mechanical. If a rewrite claims to fix quantification, it must contain a numeral the original did not have. If it does not, the label is stripped. Where the model genuinely cannot know the number, the honest output is a placeholder like [X%] for you to fill in, not a confident adverb.
Catch it yourself: search the rewrite for a digit. No digit, no quantification.
The dropped metric
This is the expensive one, because the output reads better than the input. Fluency improves and evidence disappears.
Shorter, smoother, and strictly worse. Three specific figures and the user count are gone. This is the version a hiring manager cannot act on.
Numbers are the scarcest thing on a resume. Most candidates have very few, and the ones they have are what separate their bullet from an identical bullet written by someone who did less. Trading them for readability is close to always wrong.
We reject any rewrite that loses a numeral from the original, without exception. Proper nouns get a softer rule, since a condensing rewrite can reasonably drop a descriptive aside, but numbers are absolute.
Catch it yourself: read the two versions side by side and count the figures. This is the single highest-value check on the list.
The gap you don't actually have
When a model compares a job posting against your resume, it sometimes files a requirement as missing and then, in the same breath, explains where you meet it.
Missing: Master's degree or PhD in Computer Science or a related technical field
"You have a Master of Science in Computer Science from the University of Maryland, Baltimore County, which is a strong related technical field."
That is a real example from our own product, reported by a user. The claim and its refutation shipped in the same card. Beyond looking foolish, it inflates the number of things you think you have to fix.
We now check every claimed gap against its own explanation, and move the contradicted ones to the covered side. The rule is deliberately conservative: an affirmation followed by any hedge stays a gap, because "your degrees imply a strong foundation, but you have not explicitly quantified five years of experience" is a genuine partial gap, not a misfile.
Catch it yourself: read the reasoning under each gap, not just the headline. If the explanation describes where you meet the requirement, cross it off.
Advice about text that isn't there
Resume feedback has a canon: remove personal pronouns, avoid buzzwords, cut the objective statement. Models reproduce that canon whether or not it applies to the document in front of them.
We repeatedly saw "remove personal pronouns from this section" returned for sections containing no pronouns at all, and "avoid buzzwords" aimed at words that are ordinary technical vocabulary in the candidate's field. That second one is worth dwelling on: a list of banned buzzwords that includes "framework", "scalable", or "pipeline" will flag a perfectly good engineering resume for using the words engineers use. We restrict our own buzzword list to phrases that carry no concrete meaning in any field, like "team player" and "results-driven", and check pronoun advice against whether pronouns actually appear.
Catch it yourself: before acting on a piece of generic advice, confirm the problem exists in your document. Use ctrl-F.
The formatting tell
Separate from correctness, there is recognisability. The strongest single tell in AI-written resume copy is the em dash used to splice clauses together, a construction most people rarely type and models produce constantly.
Factually fine. But recruiters have read a lot of this by now, and the punctuation alone reads as generated.
We now rewrite these automatically before a suggestion is shown, converting spliced em dashes to commas while leaving en dashes in date ranges alone. If you are working in a chat window, ask for the same thing explicitly, along with the vocabulary that clusters with it: leveraged, spearheaded, utilized, robust, seamless.
What to do about it
- Work one bullet at a time. Whole-resume rewrites produce too many simultaneous changes to audit, and that is exactly when a metric goes missing.
- Keep the original in front of you. Almost every failure above is invisible reading the output alone and obvious in a side-by-side.
- Treat numbers as immutable. If a figure did not survive the rewrite, the rewrite is wrong, however well it reads.
- Verify claims against your own memory, not the model's confidence. It has no access to your history beyond what you pasted.
- Read the reasoning, not just the verdict. Contradictions are usually visible in the explanation attached to the finding.
If you want the prompts that reduce how often these come up in the first place, they are in Claude Resume Prompts That Work. The short version: spend most of the prompt on constraints, and explicitly allow the model to answer "no honest improvement".
Get suggestions that already passed these checks
Resunova runs every AI rewrite through the validators described in this post before it reaches you. Dropped metrics, no-op edits, and false quantification claims never make it to the screen.
Score my resume freeFrequently asked questions
Does this mean AI is useless for resume writing?
No. It is genuinely good at rephrasing, condensing, and spotting where a bullet is vague. The failures cluster in a specific place: claims about your history that the source text does not support. Use it for wording and structure, and verify anything factual yourself.
Do these problems happen with every model?
The patterns we describe here are not specific to one provider. Stronger reasoning models produce cleaner output and trip the validators less often, but none of them eliminated the checks. We still run every suggestion through the same filters.
How can I check for these myself?
Put the original and the rewrite side by side. Confirm every number survived, confirm a quantification claim actually added a digit, and confirm you can say in one sentence what got better. If you cannot, keep your original line.
Why would a model list a gap the resume clearly fills?
It is extracting requirements from the posting and checking them against the resume as two separate steps, and the check is imperfect. The giveaway is that its own explanation often describes where you meet the requirement, in the same paragraph that calls it missing.