Position paper, June 2026

AI that doesn't fabricate.

Most AI resume tools optimise for novelty, so they invent. Offersly does the opposite: every suggestion is grounded in what the candidate has already written, and the model is prohibited from adding skills, metrics, or experiences the candidate did not provide. This page is the public version of that rule, the proof it works, and the answers we owe the recruiters who read what our users send.

See the boundary Read the recruiter FAQ

Written by the Offersly team. Updated as the system changes.

The principle

A resume is a record of a person's working life. Anything an AI adds that did not come from that person is a lie about a stranger's career, written by a machine, signed by them. Offersly will not write those sentences. The model rewrites, restructures, and matches phrasing to a job description using only the candidate's real material. When the source is silent on a number, our AI is silent on it too.

What our AI is allowed to do. What it cannot.

Two columns from one prompt. Drafted by engineers who wanted a rule a reviewer could check on a Tuesday morning, not a marketing slogan.

Allowed

What the model may do

Rephrase a bullet for clarity, keeping every concrete noun from the source
Match phrasing to the job description vocabulary when the meaning is unchanged
Restructure sections so the highest-relevance content sits closer to the top
Surface a missing keyword as a coaching prompt for the candidate to answer
Translate the resume between English and another listed language
Strip filler, soften corporate cliché, fix tense and grammar

Forbidden

What the model must refuse

Invent a metric. No revenue, headcount, percentage, or time saved
Add a skill, tool, or certification the candidate did not list
Insert a company, title, or employment date the candidate did not provide
Extend a duration, soften a gap, or rewrite a job into a title that was not held
Generate a bullet for an experience the candidate never wrote about
Imitate a reference, recommendation, or quoted phrase as if it came from a third party

Offersly · prompt boundary v3.1SHA · b9c4 · review weekly

Same source, three fates.

One bullet, written by a real candidate, run through three systems. The first column is what they wrote. The second is what Offersly does. The third is what most other AI tools do, and why recruiters started filtering them out.

01 · The candidate wrote

Their original bullet

Led the redesign of the seller dashboard. Worked with engineering and design weekly. Sellers said the new layout was easier to use, especially for finding their listings.

The truth: what actually happened, in the candidate's own phrasing. Often serviceable, occasionally clumsy, never inflated.

02 · Offersly tailors

Same facts, sharper sentence

Led the seller dashboard redesign with engineering and design partners, running weekly working sessions through the rebuild. Seller feedback after launch flagged listing discovery as the most improved task.

Every noun in this version traces to the bullet on the left. Verbs are stronger, structure follows the JD, no number was invented.

03 · Generic AI fabricates

What other tools shipped

Spearheaded a customer-centric overhaul of the seller dashboard, driving a 34% lift in weekly active sellers and recovering SGD 2.1M in attributed GMV across the SEA region. Managed a cross-functional pod of 12.

Three plausible-looking inventions in two sentences. A percentage, a dollar figure, and a leadership scope the candidate never claimed.

Examples are real outputs from the Offersly model on the left two panels. The fabrication panel is the consensus output of three other AI resume tools on the same input. Raw transcripts available on request.

Why this matters now, not later.

The cost of a fabrication-tolerant resume tool moved from awkward to career-ending in three steps. They happened on the calendar below.

01
2024 → 2025
Recruiter inboxes hit the spam threshold.
Median applications per role tripled. ATS systems started deduplicating by phrase. Hiring teams stopped reading second pages. The cost of a misleading bullet stopped being a polite stretch and became the reason a recruiter closes the tab.
02
2025
Reference checks went structural.
Two of the largest Singapore employers added LinkedIn-history reconciliation as a pre-interview step. A metric on a resume that does not appear anywhere in a candidate's public footprint is now a question, not a credential. Fabricated bullets stopped getting candidates further; they started getting them flagged.
03
2026
The platforms started naming names.
Recruiter tooling now surfaces probable AI-generated phrasing on the resume preview. The candidate cannot see this layer. The recruiter can. A tool that produced the fabrication is implicated by the highlight. Offersly's rewrites do not trigger these flags, because Offersly's rewrites do not invent.

The three commitments behind the claim.

Plain English on the left. The mechanism on the right. Linked to the source you can check.

Every suggestion is grounded in your source.

The model is given your existing resume text as the only acceptable raw material. It is told, in the system prompt, that anything it returns must trace to that source. If it cannot rephrase a bullet without inventing a number, it returns the bullet unchanged and explains why.

Excerpt · prompt boundary

SOURCE-ONLY: Use only the candidate's
own content. Do not invent skills,
companies, durations, or metrics.
If a stronger phrasing would require
new content, return the original
and surface a coaching prompt.

Photos and personal identifiers never reach the model.

Resume photos are stripped server-side before any AI request. Names and emails are kept (the model needs them to format), but no profile picture leaves your browser via our AI pipeline. Verified in code at functions/_lib/sanitize.js.

Excerpt · sanitiser

sanitizeForAI(resume) {
  delete resume.photoUrl
  delete resume.photo
  delete resume.avatar
  return resume
}

Every AI output is a suggestion, not a save.

No AI rewrite is committed to your resume without your explicit click. The accept step is the legal handoff: you read it, you approve it, you sign it. This is also the reason our Terms put resume accuracy responsibility on the candidate, in plain language, before signup.

Excerpt · Terms 4.3

You are solely responsible for
reviewing and editing AI-generated
content before using it in any
job application. Do not submit any
resume to an employer that contains
statements you cannot personally
verify.

Questions a recruiter actually asks.

We get these in DMs from recruiters who realised our users were reaching them. The answers below are the same ones we give privately.

You usually cannot, and we think that is the correct answer. A resume tailored with Offersly is the candidate's own content, re-ordered and rephrased. There is no watermark and no telltale sentence pattern, because the model is not the author. If you want a marker, ask the candidate to walk through any bullet in detail. An Offersly bullet will hold up because it came from their lived work.

One more thing

If you ever read a fabricated bullet on an Offersly resume, tell us.

Email a screenshot to [email protected]. We investigate every report. If our model invented something, we publish the postmortem on this page and tighten the prompt. This is the only way a claim like ours stays earned.

Try Offersly free Read the trust page

Security & disclosurePrivacy policyProduct FAQ