Tailoring a CV

The tailoring pipeline is a Claude Code skill, not a build script — the interesting parts are conversations, and two of them are approvals.

/sumisura:tailor-cv

Three ways to start

Input	What happens
A job description (pasted text or a URL)	Selection and Rewrite target that job; a cover letter is drafted too
A tracked Application (“the Acme backend role”)	Same, using the Job Description already saved, and the Generation is recorded back against that Application
Nothing (Default Mode)	A general-purpose CV from your strongest material — no cover letter

The pipeline

1. Selection — which Entries belong on this one page.
2. Rewrite — bullets re-worded toward the job’s language, your About me (summary) adapted the same way, and a row of at most ten of each Entry’s own tags picked for the job, and the headline and each employer’s role chosen from the options your profile lists. Plus a cover letter draft when there is a job, drawn from your Cover Letter Snippets if you have any.
3. Prose clean-up — the rewritten bullets and the cover letter go through the clean-user-facing-text skill (an upstream skill pinned by commit and hash, downloaded on first use): a style score, a deterministic pass that removes invisible Unicode and typographic artefacts, and a rewrite only when the score is high, under the same no-new-facts rule. You see the scores and the cleaned text at Text Review. Needs Python 3.10+ (standard library only); without it the skill stops and asks you to install it.
4. Groundedness check — every rewritten bullet compared against its source.
5. Text Review — you approve. CV and cover letter are approved independently. Nothing renders before this.
6. Render — Typst produces the PDFs.
7. PDF checks — page count, the ATS Report (see below), and the document language.
8. Visual Review — you approve the rendered result.

The prose clean-up is the one step that changes text, and it does so before Text Review, where you see the result. Checks are advisory by design: exit 0 clean, 1 flagged, 2 could not run. They inform a checkpoint where you are already looking; they never block or silently edit.

The ATS Report

Applicant tracking systems read a PDF’s text layer, not its layout. Every Generation keeps an ATS Report of what that layer says, for the CV and the cover letter:

• The extracted text, top to bottom, as a screener reads it.
• Field by field: your name, contact details (email, phone, LinkedIn, GitHub), each section header, each employer and project. Each one is found, missing or out of order. Anything missing or out of order turns the report into a warning.
• Job Description terms: which of the job’s terms that are also tags in your Master Data made it onto this CV. Only your own tags count, so the report never nudges you to claim a skill you don’t have. This part is for information and never raises a warning. It is left out in Default Mode.

The report is saved with the Generation, so you can reopen it right before you send the CV, even after output/ is cleaned up. Open it from Open the ATS Report at Visual Review, from an Application’s Generation history, or from Generated CVs. A Generation made before reports were kept says so instead of showing a pass.

Where the files go

Every run gets its own directory, so a later run never overwrites an earlier one:

output/<label>-<UTC yyyymmdd-hhmmss>/
├── cv.pdf
├── cover-letter.pdf
├── data.json          # what the template rendered
├── ats-report.json    # the ATS Report
└── selection.json     # Entries chosen + every source→rewrite pair

selection.json is what lets a groundedness verdict be re-derived later. output/ is gitignored and safe to delete; Application records survive and show a “no longer on disk” note.

Running a check by hand

./plugins/sumisura/skills/tailor-cv/scripts/quality-check.sh \
  groundedness --selection output/<slug>/selection.json

./plugins/sumisura/skills/tailor-cv/scripts/quality-check.sh \
  pdf --pdf output/<slug>/cv.pdf --data output/<slug>/data.json

Both wrap backend/cmd/cvcheck, the same Go code the web app calls. Add --json for machine-readable output. groundedness reads Master Data from data/ unless you pass --data-dir <dir>. No running backend is needed.

pdf takes a few more options for the full ATS Report:

• --job-description <file> adds Job Description term coverage. The tags come from data/, or from --data-dir <dir>.
• --cover-letter output/<slug>/cover-letter.pdf --cover-letter-data output/<slug>/cover-letter-data.json checks the cover letter too.
• --report-out output/<slug>/ats-report.json writes the report file the skill later attaches to the Generation record.

Rendering manually

typst compile --root . template/cv.typ output/<slug>/cv.pdf \
  --input data=output/<slug>/data.json

The templates are pure presentation: they read one JSON file and lay it out. Changing your CV’s look means editing template/cv.typ, not the pipeline.