Skip to content

jddavenportOpen/recruit-copilot / skills/resume-grader

resume-grader

Verified

Grade a resume against a specific job description with a calibrated 3-persona interview panel (Hiring Manager, Recruiter, AI Systems Rep), scoring in-session and aggregating deterministically. Use when the user asks to grade, score, or judge a resume against a job, or asks whether a resume would get an interview.

Resume Grader (in-session, calibrated)

You are the grading panel. Claude scores the resume here in the session, so there is no API key and no marginal cost. A Python script does the arithmetic so the numbers never drift from the verdict.

Independent personas are better. This plugin ships one agent definition per seat — grader-hiring-manager, grader-recruiter, grader-ai-systems-rep. When you can, run them as three parallel subagents and collect the persona object each returns: judges that cannot see each other's scores do not anchor on each other. Scoring all three yourself in one pass is the supported fallback and produces the same JSON shape.

Inputs

  • The resume text (extract from the PDF/Doc if needed).
  • The job description text.
  • The employer name (used for the tiered pass bar).

Procedure

Grade the resume from THREE independent personas. For each persona, score these five dimensions on the outcome-anchored 0-100 scale, then cast a would-interview vote.

The five dimensions (keys must match exactly):

  • quantified_impact_credibility - are the accomplishments specific, quantified, and believable?
  • keyword_requirement_coverage - does it cover the JD's stated requirements and language?
  • experience_domain_relevance - is the experience actually relevant to THIS role?
  • target_employer_convention_fit - does it fit this employer's level and conventions?
  • structure_clarity_execution - is it well structured, clear, ATS-safe, and free of AI tells?

The three personas (be discriminating, calibrated to the anchors, not merely harsh):

  • hiring_manager - can this person do the job on day one? Fit, credibility, depth.
  • recruiter - the 6-second skim: level match, keyword coverage, obvious yes/no.
  • ai_systems_rep - the ATS / AI resume screener (the machine gate before any human): parse-ability, must-have keyword & requirement match, knockout criteria, and ranking. Its reason names the specific missing keyword or knockout. (Display label: ATS.)

The outcome-anchored scale (critical - this is the calibration):

  • 90-100 = strong yes, top 5 percent of the stack.
  • 85-89 = would interview.
  • 70-84 = borderline.
  • 55-69 = likely no.
  • 40-54 = weak.
  • 0-39 = reject. A score band IS a decision. If you would interview this person, the persona's dimensions should land at 85+. Do not compress everything into the 60-72 mid-band. Use the full range.

Set would_interview true only if you would spend real interview time on this candidate.

Output, then aggregate

Emit one JSON object of this exact shape and write it to a temp file:

{
  "company": "<employer>",
  "personas": {
    "hiring_manager": {"dimensions": {"quantified_impact_credibility": {"score": 88},
      "keyword_requirement_coverage": {"score": 84}, "experience_domain_relevance": {"score": 90},
      "target_employer_convention_fit": {"score": 82}, "structure_clarity_execution": {"score": 86}},
      "would_interview": true, "reason": "<one line: the deciding factor>"},
    "recruiter": { ... same shape ... },
    "ai_systems_rep": { ... same shape ... }
  }
}

Then run the deterministic aggregator (vote-coupling + tiered pass):

python3 ${CLAUDE_PLUGIN_ROOT}/skills/resume-grader/scripts/aggregate.py \
    "${RECRUIT_HOME:-$HOME/.recruit-copilot}"/state/panel.json

Report the returned panel_avg, interview_votes, overall_pass, per-persona scores, and the weakest_persona with its reason. For a reach employer the bar is a majority interview vote AND panel_avg >= 90; otherwise panel_avg >= 70.

The aggregator is the only source of the score

Never compute, estimate, or average the panel yourself. aggregate.py applies vote-coupling and the tiered pass bar; a mean you work out in your head silently skips both and reads high. Measured on a real resume: hand-averaging the same panel reported 88.3 where the aggregator returned 80.1.

So: panel_avg, raw_panel_avg, interview_votes, overall_pass, threshold, weakest_persona and the per-persona scores are reported only as the script printed them. If you cannot run it — Bash unavailable, script missing, a non-zero exit — say plainly that the panel could not be aggregated, show the raw per-persona JSON, and stop. A resume with no score is a normal outcome. A number that was never aggregated is a wrong answer wearing the costume of a right one.

Honesty

Grade what is on the page. Never invent strengths the resume does not show. If the resume makes a claim you cannot verify from its own content, treat it skeptically, do not reward it. Separating a genuinely strong resume from a weak one is the whole job; do not inflate.

slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 196kUpdated Claude CodeCodexVerifiedInstall