leansii/vouch-skills

**Job-tailored resumes where every line is something you can vouch for.**
AI resume tools have one failure mode in common: they write a better candidate
than you are. A Kafka you never ran, a fintech you never worked in, "built" where
you led the people who built it. It reads fine, it passes the ATS, and it falls
apart in the first interview.
Vouch is an Agent Skill for Claude Code, claude.ai and Antigravity (Codex and ChatGPT
adapters included, untested)
that tailors your CV and cover letter to a job posting **only from facts you've
recorded**, then checks every line in a separate pass that never saw the draft
being written.
Real output on the fictional example in examples/; the verdicts came from
DeepSeek V4 Flash judging the packet.
The ATS happily counts the invented Kafka as a match. The verification step is
what catches it.
How it works
1. **Corpus.** Your experience as small Markdown files: one per employer, one
record per achievement, every number tagged by how sure you are —
verifiable, estimate, from-cv, cannot-confirm. The skill interviews you
to fill it in and to tag it honestly.
2. **Tailor.** The CV and letter use only records selected for this posting.
Provenance decides wording: estimates are never printed as precise figures,
unconfirmed numbers never at all. Attribution stays exact. No gap apologies,
no recasting your e-commerce work as fintech.
3. **Verify.** Code extracts every claim and the evidence; a judge in a clean
context (a sub-agent where your tool has them) answers "supported?" per line;
code maps each verdict through the provenance policy to keep / soften / flag /
remove. A missing or unreadable verdict is *unverified*, never a silent pass.
4. **ATS check.** Parses the rendered PDF/DOCX the way applicant-tracking systems
do, finds the posting's must-haves, and splits what's missing into **dropped
facts** (you have them, the CV didn't say them — fix) and **true gaps** (you
don't — leave them out). The score can only go up honestly.
5. **Before and after.** *Fit* checks a posting first — still open? anything
that rules you out (no visa sponsorship, must be based in X)? how much of it
your corpus honestly covers? *Stories* turns records into STAR interview
stories that code checks against the same records.
6. **You send.** Nothing is submitted, emailed or posted. Ever.
The decisions — what counts as a claim, what evidence the judge sees, what a
verdict means — live in dependency-free Python (skills/vouch/scripts/), not in
a prompt. Any model can do the writing and judging.
Install
| Tool | How | Tested |
|---|---|---|
| **Claude Code** | /plugin marketplace add leansii/vouch-skills then /plugin install vouch@vouch | ✅ end to end; the judge runs as a separate vouch-judge subagent |
| **claude.ai** | Seven single-file skills in adapters/claude-ai: Settings → Customize → Skills → Add → Upload a skill, select the seven .md files. | ✅ verify, ATS, tailor; the judge runs in the same chat |