candidate-source-audit

Resume is packaging; code is truth. 简历是包装,代码是真相。

A Claude skill that turns "evaluating an engineering / AI job candidate" into a repeatable, source-level due-diligence SOP: instead of trusting the résumé, it clones the candidate's GitHub, reads the actual source code, verifies who really wrote it (via git history), checks the work against a role's anti-indicator list, tells a real system apart from a thin wrapper ("套壳"), assesses genuinely AI-Native thinking, weighs potential & retention by year-of-study × major, and produces an evidence-backed ranking — plus interview questions tailored to the candidate's real projects, a one-page scorecard, and a pre-screen script.

Why

Anyone can polish a résumé. Reading the résumé and judging from it is letting the candidate grade themselves. This skill shifts the question from "what do they claim" to "which claim survives a look at the source, and which one collapses on inspection."

What's inside

candidate-source-audit/
├── SKILL.md                       # 8-step SOP + quick tells + output format
├── scripts/
│   ├── find_github.py             # locate the real account (dead links, renames, email-derived handles, variants)
│   └── github_enum.py             # enumerate repos + telltale metadata (fork ratio, size, recency)
└── references/
    ├── rubric.md                  # the judgment core: dimensions · shell-vs-system tells · anti-indicators · AI-Native · potential · "unverifiable ≠ fabricated"
    ├── role-bar.md                # the role's bar — a TEMPLATE you customize per role
    └── interview-and-prescreen.md # interview-question design · one-page scorecard · informal pre-screen script

Install (Claude Code)

Drop the folder into your skills directory:

git clone https://github.com/Zion74/candidate-source-audit.git \
  ~/.claude/skills/candidate-source-audit

It then auto-triggers whenever you ask Claude to screen / evaluate / vet / rank a candidate, or paste a résumé + GitHub. You can also invoke it explicitly with /candidate-source-audit.

Customize for your role

Edit references/role-bar.md — that's the yardstick. Put your role's hard requirements and anti-indicators there; the audit's "fit" and "anti-indicator" judgments follow from it. The shipped example is a generic "AI Agent Harness" engineering role.

Standalone scripts

The two scripts work on their own (no auth needed; GitHub's unauthenticated API allows ~60 req/hr):

python scripts/find_github.py octocat octocat@example.com "The Octocat"
python scripts/github_enum.py <github-handle>

Responsible use

This skill evaluates people. Use it only for legitimate hiring/evaluation, only on publicly available material, and keep conclusions evidence-based and fair. "Can't find it" is not proof of fabrication — the rubric explicitly separates private/unfinished from exaggerated/nonexistent (absence of evidence ≠ evidence of absence). The human, not the AI, owns the final call.

License

MIT