Selection, not generation

Your resume should only say things that are true.

Ask an AI to tailor your resume and it will write you a better one — including the parts you never did. Lapidly works the other way round. You approve the evidence once; tailoring can only ever select from it.

No credit card. Your evidence bank stays yours.

Why this exists

“Do not fabricate” is a request.

Lapidly was built after auditing 21 applications produced by a conventional AI resume tailorer — one that received the resume and the job posting, and was told to tailor it without inventing anything. It complied most of the time. Here is what most of the time looked like:

7

technologies invented

Each appeared in exactly the one resume sent to the company whose posting asked for it. That pattern survives a keyword filter and dies in the first technical screen.

3

resumes naming real customers

Client names the candidate's employment contract prohibits disclosing. A contract problem, not a writing problem.

74–91%

match scores that predicted nothing

Generated by the same model that wrote the resume. A number grading its own homework is decoration.

None of these are prompting failures, and no amount of prompt tuning fixes them. They are what an open-ended generation step produces when the only thing standing between it and the page is an instruction. The fix has to be structural.

The method

A cutter never adds material.

Lapidary work is subtractive. You start with the stone you have and remove everything that is not the finished gem. That is the entire architecture, and it is why the failure modes above cannot occur here rather than merely being discouraged.

  1. 01

    Build your evidence bank once

    Every achievement, metric, and skill you can defend in an interview — written by you, approved by you. Locked fields like titles, employers, and dates are byte-identical on every document forever.

  2. 02

    Get an honest read before you apply

    Paste a posting and Lapidly maps each requirement to your evidence: strong, partial, or absent. Coverage is a count — meets 7 of 7 — not a percentage. Sometimes the answer is that you should skip this one.

  3. 03

    Tailoring selects; it cannot write

    The model chooses which of your approved bullets to use and in what order. It returns identifiers, never sentences. There is no channel through which a new claim can reach the page.

  4. 04

    Nothing ships until it passes the gate

    The finished document is rendered, read back by two independent parsers, and checked against every rule. A blocked document is not deliverable. There is no override.

The gate

Fails closed.

A resume can look perfect on screen and extract as nonsense. One real example: letter-spacing on section headings rendered beautifully and extracted as EDUCA TION. No preview would have shown it. No match score would have flagged it.

So every document is read back the way a parser reads it, by two engines that share no code. If they disagree about a single word, that is a failure — because the applicant tracking system on the other end is a third parser nobody can inspect.

PARSE GATEPASS — safe to send
  • Customer confidentiality

    No client name reaches a document.

  • Verified skills only

    Anything outside your inventory is a gap, not a bullet.

  • Locked identity

    Name, contact, titles, and dates byte-identical.

  • Metric canon

    One figure per accomplishment; no stacking.

  • Reading order

    Your name is the first text a parser sees.

  • Two-page cap

    Enforced by trimming the weakest evidence.

  • Intact keywords

    Catches spacing that splits words on extraction.

  • Cross-parser agreement

    Two independent engines must read the same text.

Fewer applications. Better ones.

A gate that occasionally tells you to skip a posting is worth more than a generator that never does. Twenty-one applications and almost no interviews is what the fast path produces.

Questions? How Lapidly came about