An AI resume builder that can't make things up

Why do AI resume builders invent experience?

Not because they are badly built. Because of what they are.

A language model asked to make your resume match a job description is being given two instructions that quietly conflict. One is "match this job." The other is "stay true to this person." When those two things can both be satisfied, the model rephrases your real work to speak the job's language, which is genuinely useful. When they cannot, something has to give, and the model has no way to know which instruction is the important one.

So it fills the gap. It reaches for the Kubernetes experience the job wants and you do not have. It rounds "helped with" up to "led." It attaches a number to an achievement that never had one. The output is fluent, confident and completely plausible, because fluent and plausible is exactly what the model was trained to produce. It is not lying in any way it could detect. It ran out of facts and kept writing.

An AI does not fabricate because it is dishonest. It fabricates when it is under pressure to produce and has run out of things it actually knows.

That framing points at the fix. If fabrication happens when the model runs out of facts, then you solve it by giving the model more facts and removing its permission to write without them. Not by asking it nicely in a prompt.

What is the Calibration Engine?

The Calibration Engine is the set of checks a tailored resume has to pass before JobShifu will produce a document. A tailored resume is treated as a build artifact, the way compiled software is: it either passes and exists, or fails and does not. The product enforces that, not your good intentions.

One check blocks. Seven report.

The split is deliberate, and it is worth being precise about which check carries the guarantee.

Provenance is exact and deterministic. It is ordinary code reading the citations attached to each line and confirming they point at confirmed Vault units. It cannot be argued with, it has no opinion, and it returns the same verdict every time. That is what makes it safe to block on.

Scope and entailment use a model, so they advise rather than block. Judging whether a line overstates its source is a judgement call, and a model making judgement calls will sometimes be wrong. Blocking your resume on a check that can be wrong would be its own kind of dishonesty. So they flag, you decide, and the deterministic check is the one holding the floor.

How does tailoring actually work?

The short version: JobShifu proposes, you approve, and the gate runs last.

1. Your Vault holds confirmed facts

   Roles, projects, achievements, skills. You can import a resume to start it, but nothing becomes citable until you confirm it. Unconfirmed material is invisible to the tailoring step, which means the model cannot build on something you have not stood behind.

   The Vault also grows as you read: a job description works like a memory prompt, surfacing wins you forgot or never thought employers cared about, and confirming one with a tap makes it citable for every resume after.

2. The model proposes changes, it does not write your resume

   JobShifu reads the job description and returns a set of specific edits: keep this line, rewrite that one, drop this, combine those two, add one built from these Vault entries. Every proposal names the confirmed units it draws from. Proposals citing anything unconfirmed are dropped before they reach you.

3. You approve every change, card by card

   Each proposal is a card showing the original, the proposal, and the evidence behind it. Accept, edit, or reject. There is no "accept all" that skips the reading, and that is on purpose: the review is where you catch the line you would not want to defend, and it is also how JobShifu learns what you actually sound like.

4. The gate runs before the document exists

   Your approved changes are assembled and the full Calibration Engine runs on the result. Provenance passes or the build fails. Only on a pass does a finished resume version exist to preview, download or attach.

What happens when a resume fails the check?

You get told which line failed and why, and no document is produced.

This is the part people find surprising, so it is worth stating plainly: a failed build is not a degraded resume with a warning on it. It is the absence of a resume. There is nothing to accidentally download, nothing to forget you were warned about, nothing sitting in your documents folder six weeks later that you attach to an application without remembering the banner you dismissed.

A warning you can click past is a warning that eventually gets clicked past, usually at 11pm on the night before a deadline. That is exactly when you least want the safety net to be optional.

Does this mean my resume gets worse?

It means your resume stops borrowing from experience you do not have, which is a different thing.

Most of what makes a tailored resume work is not invention. It is selection and language: choosing which of your real achievements to lead with for this specific job, and describing them in the terms this employer uses. Both of those are fully available inside the constraint. JobShifu will happily rewrite "reduced page load time" as "improved frontend performance" when the job description asks for the second phrasing, because that is your real work in the job's vocabulary.

What it will not do is give you the distributed systems experience the job wants and your Vault does not contain. And here the constraint is doing you a favour, because that line was never going to survive the interview anyway.

There is also a compounding effect that only works if the facts are real. Every job you read teaches JobShifu something about what employers keep asking for, and the gaps it finds turn into concrete things you can go and actually do. A Vault built from real work gets genuinely stronger over time. A resume built from invention has to be maintained as a story you remember.

Frequently asked questions