ai-resume-tuner

ResumeHQ — the resume tool that refuses to lie for you

Every AI resume tool promises to "beat the ATS." This one has a harder rule: it never invents experience — and when a job is a genuine mismatch, it declines to tailor at all and tells you why. Finds jobs, scores your fit with deterministic engines (not LLM vibes), tailors through a fail-closed pipeline where an independent auditor can veto the writer, and publishes its own scoring failures. Works as a Claude Code plugin, Codex plugin, or standalone web app.

Upload a resume → paste the job posting → keyword-match and recruiter-style heuristic scores with concrete fixes, in under 30 seconds. Try it free.

---

Why ResumeHQ?

Most resume tools only score the resume you bring to them. ResumeHQ goes further:

Feature	Jobscan	Rezi	Teal	ResumeHQ
Keyword + semantic match scoring	✅	✅	✅	✅
Recruiter-style heuristic review	❌	❌	❌	✅
Discover matching jobs	❌	❌	❌	✅
Score jobs against your resume	❌	❌	❌	✅
Auto-tailor resume to JD	✅	✅	❌	✅
ATS-compliant DOCX output	❌	✅	❌	✅
Application tracker	❌	❌	✅	✅
Works in Claude Code / claude.ai / Codex	❌	❌	❌	✅
Open source	❌	❌	❌	✅
Publishes its scoring benchmarks — including the bugs	❌	❌	❌	✅
Refuses to fabricate — declines to tailor genuine mismatches, audits every claim against your real resume	❌	❌	❌	✅

---

What This Does

You paste a job description (or search for jobs). The system:

1. Discovers matching jobs from live job boards — scored and ranked by fit with your resume
2. Gates candidate fit — your master resume must clear the fit bar (default 50) against the exact JD with zero hard knockouts before resume work begins. If you don't fit, it says so and stops — no tool that invents experience to close the gap is working for you
3. Analyzes passing JDs — extracts keywords, required skills, domain, seniority level
4. Tailors your master resume — rewrites bullets, reorders sections, matches terminology
5. Scores the result with two independent advisory heuristics (keyword/semantic match + recruiter-style review)
6. Iterates automatically until the evidence match covers every must-have (75%+ evidence match) — an internal editorial target, not a prediction of any employer's screening
7. Generates production-ready DOCX files (resume + cover letter)
8. Tracks every application in an Excel spreadsheet

The candidate-fit gate always runs first and cannot be bypassed by ATS/HR scores. After it passes, safe read/scoring work may run concurrently while authorization, DOCX generation, and tracker mutation remain ordered.

---

Quick Start — 3 Steps

Works with Claude Code (CLI/IDE), Codex (CLI/app/IDE), and claude.ai (web/Projects).

Step 1: Install the plugin

Claude Code:

/plugin marketplace add jananthan30/Resume-Builder
/plugin install resume-builder

Codex from a local checkout:

codex plugin marketplace add .

Then restart Codex and install Resume Builder from the Resume Builder Local marketplace.

Step 2: Configure the runtime

Claude Code exposes the plugin setup command:

/resume-builder:setup

This walks you through everything:

• Checks if Python is installed (tells you where to download it if not)
• Installs all dependencies automatically (pip install -r requirements.txt)
• Creates your config.json with your name, email, phone, LinkedIn
• Optionally links a Pro account for unlimited cloud scoring
• Optionally sets up the LLM scorer (Claude API key)

For Codex, install Python 3.10+, run python -m pip install -r requirements.txt, and create config.json with a valid master_resume_path. The installed Codex surface exposes the Resume Team as $resume-team; it does not expose the Claude-style /resume-builder:* command namespace.

Step 3: Start building resumes

Claude Code:

/resume-builder:resume [paste a job description here]

Codex:

$resume-team [paste a job description here]

$resume-team publishes an authorized, digest-verified resume.md draft. It does not by itself create a DOCX or complete an application package.

Or find jobs first:

/resume-builder:find-jobs Senior Data Scientist in New York

---

Claude Code Slash Commands (9)

Command	What It Does
/resume-builder:setup	One-time setup wizard (installs Python deps, creates config, links Pro account)
/resume-builder:job-fit [JD]	Deterministic master-vs-JD gate (fit bar, default 50, and zero hard knockouts) before tailoring
/resume-builder:resume [JD]	Full application: tailored resume + cover letter + scoring + DOCX + tracking
/resume-builder:tailor-resume [JD]	Resume only (no cover letter)
/resume-builder:cover-letter [JD]	Cover letter only
/resume-builder:find-jobs [title] [location]	Discover and score matching jobs from live job boards
/resume-builder:batch-resume	Process multiple job descriptions in parallel
/resume-builder:writing-coach [file]	Audit and rewrite resume bullets using 10 writing rules
/resume-builder:resume-team [JD]	Publish an authorized resume.md draft through the native Researcher → Writer → Auditor → Editor workflow

If running Claude Code locally from the cloned repo, use short names: /resume, /tailor-resume, /find-jobs, etc. In Codex, invoke $resume-team.

What Works Without Setup

Some Claude Code commands can provide prompt-only previews before setup. Production resume generation through /resume-builder:resume, /resume-builder:resume-team, or Codex $resume-team requires Python, config.json, the deterministic candidate-fit preflight, and the evidence, human-voice, and canonical-integrity audit helpers; those gates are never skipped.

Command	Works immediately?	With setup?
/resume-builder:job-fit	No — requires the configured master and deterministic preflight	Digest-bound score, threshold, and hard-knockout decision
/resume-builder:resume	No — the native team and deterministic audits require setup	Full audited resume + automated ATS/HR scoring and DOCX output
/resume-builder:resume-team / $resume-team	No — requires macOS/Linux, the configured master resume, and Python audit helpers	Authorized, digest-verified resume.md draft; DOCX/tracker finalization is still pending
/resume-builder:cover-letter	Yes — the assistant writes the letter	+ DOCX output
/resume-builder:writing-coach	Yes — full writing audit	Same
/resume-builder:find-jobs	Yes — shows results (no score)	+ ATS/HR fit scoring per job
/resume-builder:setup	Yes — runs the setup wizard	N/A
MCP scoring tools	No — needs Python	evidence_match, score_resume, score_ats, score_hr, score_with_llm, explain_score, extract_text, discover_jobs

---

MCP Tools (7 production-supported surfaces)

After running /resume-builder:setup, the MCP scorer auto-starts and provides these tools that Claude Code or Codex can call natively:

Tool	What It Does
evidence_match	Requirement-by-requirement evidence matching, with the exact excerpt behind each conclusion (recommended)
score_resume	Legacy ATS + HR analysis in one call
score_ats	Keyword + semantic match scoring (8 heuristic components)
score_hr	Recruiter-style heuristic review (6 factors + F-pattern)
score_with_llm	LLM-augmented rubric scoring (requires ANTHROPIC_API_KEY)
explain_score	Actionable improvement suggestions with missing keywords
extract_text	Extract text from DOCX/PDF/MD/TXT files
discover_jobs	Search live job boards and score each job against your resume

All listed MCP tools support cloud-first scoring — they try the cloud API first and fall back to local scoring automatically. Legacy direct rewrite endpoints or functions are not production-authorized tailoring paths. The capability-isolated native Resume Team is the sole production rewrite and draft-publication path.

---

Job Discovery

The /find-jobs command and discover_jobs MCP tool search live job boards and rank results by how well each job matches your resume — answering "which jobs should I actually apply to?" with data.

/resume-builder:find-jobs Senior Product Manager in San Francisco
/resume-builder:find-jobs Data Scientist remote

How it works:

1. Searches Adzuna (16 countries, salary data) + Remotive (remote jobs) + JSearch (aggregated boards incl. niche career centers; optional RapidAPI key)
2. Pre-filters top 20 results by title relevance
3. Lightweight scores all 20 candidates (keyword + phrase + BM25 — fast)
4. Full ATS + HR scores top 10 finalists
5. Returns ranked list with scores, salary range, and apply links

Sample output:

Rank  Title                        Company        ATS   HR    Salary
────  ───────────────────────────  ─────────────  ────  ────  ──────────────
#1    Senior Data Scientist        Pfizer          82%   74%  $120k–$150k
#2    Data Scientist II            Goldman Sachs   79%   71%  $110k–$140k
#3    ML Engineer – NLP            Microsoft       74%   68%  $130k–$160k

API keys required for job search:

• Adzuna (free): Register at developer.adzuna.com — add ADZUNA_APP_ID and ADZUNA_APP_KEY to your .env
• Remotive: No key needed (remote jobs only, included automatically)

---

Evidence Matching

A universal "ATS score" does not exist. Recruiting systems parse, search, filter, evaluate requirements and sometimes rank — using materially different mechanisms — so a single percentage cannot describe them all.

This tool answers a question that is answerable:

Your resume provides strong evidence for 17 of 21 important requirements. One must-have has no evidence found. Eligibility is UNVERIFIED.

Every conclusion cites the exact resume excerpt behind it, with character offsets you can verify.

python evidence_match.py --resume resume.docx --jd job.txt --verify

Three results stay separate and are never multiplied together:

Output	Meaning
Eligibility	PASS / FAIL / UNVERIFIED on explicit hard requirements
Qualification Evidence Fit	How strongly the resume supports the role's requirements
Evidence Quality	How explicit and traceable that evidence is

A missing licence is UNVERIFIED, not FAIL — the resume did not establish it, which is not proof the candidate lacks it. Only explicit contradictory evidence (an expired licence) produces FAIL.

Why this resists gaming. The model classifies evidence; Python computes every number. Repeating a phrase twenty times collapses to one piece of evidence, a parent skill never satisfies a child requirement (Python does not imply TensorFlow), and unapproved synonyms are downgraded rather than accepted. Scores replay exactly from stored judgments with no LLM call:

python evidence_match.py --replay match.json