English · 繁體中文 · 简体中文 · 日本語

vibe-resume

Turn your AI coding history into a versioned, reviewer-ready résumé — for the vibe coding era.

vibe-resume scans every AI assistant you use on macOS (Claude Code, Cursor, GitHub Copilot, Cline, Continue, Aider, Windsurf, Zed AI, ChatGPT / Claude.ai / Gemini / Grok / Perplexity / Mistral exports, ComfyUI, Midjourney, Suno, ElevenLabs, and your git commits) and synthesizes the usage trail into a Markdown / DOCX / PDF résumé with built-in git snapshots so every draft is diff-able and rollback-able.

How it differs

	vibe-resume	Reactive Resume / OpenResume	Resume-LM / Resume Matcher	HackMyResume / JSON Resume
Primary signal	AI tool sessions + git commits (auto-extracted)	User-typed content in WYSIWYG	Uploaded PDF + JD	User-typed JSON
Locales	10 (en_US/en_EU/en_GB/zh_TW/zh_HK/zh_CN/ja_JP/ko_KR/de_DE/fr_FR) with culture-specific layouts	1–2	1	Theme-dependent
JP 履歴書 JIS Z 8303 grid	✅ render/japan.py	❌	❌	❌
Europass labelled personal-info	✅ en_EU template	❌	❌	❌
Reviewer audit	8-point scorecard + trend sparkline	—	ATS score only	—
JD tailoring	enrich --tailor JD.txt (LLM prompt injection)	—	✅ LLM rewrite	—
Privacy	Fully local; claude -p headless; nothing leaves your machine	Varies (OpenAI-key optional)	Cloud API required	Fully local
Shape	Python CLI pipeline	Web UI	Web UI	Node CLI
Agent-Skill hosts	8 (Claude Code · Gemini CLI · Copilot CLI · Cursor · Warp · OpenClaw · OpenCode · Hermes) — single canonical SKILL.md	—	—	—

Why

Hiring in 2026 rewards engineers who can prove AI-assisted productivity with measurable outcomes — not just list "Claude Code" as a skill. Reviewers want to see architecture decisions, cross-stack breadth (frontend / backend / DevOps / bug-fix / deployment), and how fast you ship. Your AI tools already log this automatically. vibe-resume turns that exhaust into evidence.

Features

Local extractors (no login required)

Source	Where
Claude Code	~/.claude/projects/**/*.jsonl
Claude Code Archive	~/ClaudeCodeArchive/current (optional rsync backup)
Cursor	~/Library/Application Support/Cursor/User/**/state.vscdb
GitHub Copilot (VS Code)	workspaceStorage/**/chatSessions/
Cline	globalStorage/saoudrizwan.claude-dev/ or ~/.cline/data/
Continue.dev	~/.continue/sessions/
Aider	$HOME/**/.aider.chat.history.md
Windsurf / Cascade	~/.codeium/windsurf/cascade/
Zed AI	~/.local/share/zed/threads/
Claude Desktop	MCP config + extensions
Git commits	every .git in $HOME filtered by your author email

Cloud export importers (drop a ZIP into data/imports/<tool>/)

ChatGPT · Claude.ai · Gemini Takeout · Grok · Perplexity · Mistral Le Chat · Poe

AIGC extractors

image_local (ComfyUI / A1111 PNG metadata) · midjourney (IPTC/XMP) · elevenlabs (history API) · suno (local MP3 ID3) · runway / heygen (stub)

Resume intelligence

• Task-type classifier — tags each session as frontend / backend / bug-fix / deployment / refactor / testing / etc.
• Capability breadth — counts distinct task categories per project; surfaces multi-skill engineers
• 30-day rolling stats — active-day ratio, daily avg, peak day, longest streak (aligned with Claude Code's 30-day cleanup)
• XYZ enricher — uses Claude Code CLI headlessly to turn raw activity into Google-style résumé bullets
• Tech stack canonicalization — postgres → PostgreSQL, tailwind → Tailwind CSS
• Domain-tag vs hard-skill separation — keeps ATS keywords clean
• Privacy filter — regex redaction + project blocklist + optional tech abstraction
• Versioned output — internal git repo under data/resume_history/ with list-versions / diff v1 v2 / rollback

Use as an Agent Skill (Claude Code · Gemini CLI · Copilot CLI · Cursor · Warp · OpenClaw · OpenCode)

Beyond the CLI, vibe-resume ships as an Agent Skill that drives the full extract → aggregate → enrich → render → review → trend pipeline from natural-language prompts. Inside any supported host you can just say:

"Generate my résumé from my AI usage." "Render my résumé in Japanese and German and review both." "Tailor my résumé for this JD: data/imports/lumen_labs.txt."

The skill follows the 2026 converged SKILL.md convention: one canonical file at skills/ai-used-resume/SKILL.md; every other host path (.claude/skills/, .gemini/skills/, .agents/skills/, .opencode/skills/) is a symlink pointing at it. For marketplace installs, .claude-plugin/plugin.json and .codex-plugin/plugin.json wrap the skill as a Claude Code / OpenAI Codex plugin.

Host	Discovery	Setup in this repo
Claude Code	.claude/skills/	Canonical — auto-loaded
Gemini CLI (Google)	.gemini/skills/	Symlink → canonical
GitHub Copilot CLI	reads .claude/skills/ natively	Zero config
Cursor CLI	reads AGENTS.md + .cursor/rules/	AGENTS.md points at SKILL.md
Warp (agentic terminal)	reads .claude/skills/ + .agents/skills/ + .warp/skills/	Zero config (.agents/skills/ symlink added defensively)
OpenClaw (250k⭐ in 60 days)	~/.openclaw/skills/ (user scope only)	User-scope symlink — see below
OpenCode (terminal CLI agent)	.opencode/skills/ + ~/.opencode/skills/	Project-scope symlink included
Hermes Agent (Nous Research)	repo skills/<name>/SKILL.md → installed to ~/.hermes/skills/<category>/<name>/	Native skill at skills/ai-used-resume/SKILL.md; hermes skills tap add easyvibecoding/vibe-resume && hermes skills install easyvibecoding/vibe-resume/ai-used-resume

Install — three ecosystem tiers

The 2026 agent-skills ecosystem has converged into four install paths — pick the one that matches your agent, not eight separate ln -s commands.

Tier 0 — Plugin marketplaces (bundled install, namespaced skill)

# Claude Code (uses .claude-plugin/plugin.json)
/plugin marketplace add easyvibecoding/vibe-resume
/plugin install vibe-resume@vibe-resume

# OpenAI Codex ≥ 0.121 (uses .codex-plugin/{plugin,marketplace}.json)
codex plugin marketplace add easyvibecoding/vibe-resume
codex
/plugins   # in the TUI: find vibe-resume and install

Codex plugin support shipped in codex-cli v0.121.0 (2026-04-15). On older Codex, fall back to Tier 1 (npx skills add …). In the TUI's /plugins pane you can also add the marketplace interactively instead of the shell command.

Plugins wrap the skill plus future commands/agents/hooks/MCP servers in one installable unit. Skill is exposed as /vibe-resume:ai-used-resume (namespaced) instead of bare /ai-used-resume.

Tier 1 — 37+ agentskills.io-standard hosts (one line installs to all)

npx skills add easyvibecoding/vibe-resume --skill ai-used-resume

npx skills auto-detects every installed CLI/IDE agent on your machine and routes the skill to the correct directory. This one command covers Claude Code, Cursor, Windsurf, Gemini CLI, GitHub Copilot, Codex, Qwen Code, Kimi Code, Roo Code, Kilo Code, Goose, Trae, OpenCode, Amp, Antigravity, Kiro, Factory, Junie, and more. To restrict, pass -a <slug>:

npx skills add easyvibecoding/vibe-resume -a claude-code -a cursor -a windsurf

Full list of Tier-1 agent slugs (for -a flag)

Slugs verified against skills@latest on 2026-04-22. Run npx skills add … -a unknown to have the CLI print the up-to-date list.

Agent	slug		Agent	slug
Amp	amp		Kilo Code	kilo
Antigravity	antigravity		Kimi Code CLI	kimi-cli
Augment	augment		Kiro CLI	kiro-cli
Claude Code	claude-code		Mistral Vibe	mistral-vibe
Cline	cline		opencode	opencode
CodeBuddy	codebuddy		OpenHands	openhands
Codex CLI	codex		Pi	pi
Continue	continue		Qoder	qoder
Cursor	cursor		Qwen Code	qwen-code
Gemini CLI	gemini-cli		Replit	replit
GitHub Copilot	github-copilot		Roo	roo
Goose	goose		Trae	trae / trae-cn
IBM Bob	bob		Warp	warp
iFlow CLI	iflow-cli		Windsurf	windsurf
Junie	junie		Zencoder	zencoder

Universal fallback: -a universal drops the skill under .agents/skills/, which Claude Code / Codex / Gemini CLI / Cursor / Warp / Amp / Cline / OpenCode / Antigravity read natively — one install covers all of them. Latest slug list: vercel-labs/skills.

Tier 2 — OpenClaw (own ClawHub marketplace + 5,400+ skill registry)

openclaw skills install easyvibecoding/vibe-resume/ai-used-resume

Tier 3 — Hermes Agent (own skills.sh registry + native 5-section body format)

hermes skills tap add easyvibecoding/vibe-resume
hermes skills install easyvibecoding/vibe-resume/ai-used-resume --force --yes

Manual install / symlink fallback (no Node, custom paths, Windows)

If you can't run npx skills or want full control over symlink locations:

# Tier 1 hosts — symlink from this repo's canonical SKILL.md (at skills/)
mkdir -p ~/.claude/skills && ln -s "$(pwd)/skills/ai-used-resume" ~/.claude/skills/ai-used-resume
mkdir -p ~/.gemini/skills && ln -s "$(pwd)/skills/ai-used-resume" ~/.gemini/skills/ai-used-resume
mkdir -p ~/.warp/skills && ln -s "$(pwd)/skills/ai-used-resume" ~/.warp/skills/ai-used-resume
mkdir -p ~/.opencode/skills && ln -s "$(pwd)/skills/ai-used-resume" ~/.opencode/skills/ai-used-resume

# Cursor reads project-root AGENTS.md zero-config. For system-wide, copy to ~/.cursor/rules/.

Windows (elevated PowerShell):

New-Item -ItemType SymbolicLink -Path $HOME\.claude\skills\ai-used-resume `
  -Value (Resolve-Path .claude\skills\ai-used-resume)
# repeat for .gemini / .warp / .opencode

How to trigger the skill once installed

In every 2026 host, the skill auto-fires when your prompt matches the description line of the SKILL.md frontmatter — so natural-language trigger phrases like "generate my résumé from my AI usage", "render my résumé in Japanese", "tailor my résumé for this JD", "score my résumé", or "show my résumé trend" are enough in most tools. Most hosts also expose an explicit invocation path:

Host	Auto-trigger	Explicit invocation
Claude Code	✅ via description match	/ai-used-resume slash command
Gemini CLI	✅ activate_skill tool loads on match	after install, run /agents refresh once to index
GitHub Copilot CLI	✅ description match	gh skill install easyvibecoding/vibe-resume (uses repo skills/ layout)
Cursor CLI	✅ AGENTS.md at project root auto-applies	content also copyable into .cursor/rules/
Warp	✅ agent picks from available skills	/ai-used-resume slash command or searchable skills menu
OpenClaw	✅ matches description at load	/ai-used-resume or openclaw skills install easyvibecoding/vibe-resume
OpenCode	✅ via internal SkillTool	/ai-used-resume slash command
Hermes Agent	✅ matches description	hermes chat -s ai-used-resume -q "generate my résumé" preload form

A quick verification prompt for any host: "Walk me through generating my résumé from AI usage — don't run anything, just describe the 6 commands in order." If the skill is correctly loaded, the response names extract → aggregate → enrich → render → review → trend and matches uv run vibe-resume syntax. End-to-end verified in Hermes with hermes chat -Q -s ai-used-resume.

See AGENTS.md for the full matrix, Windows symlink commands, and Hermes Agent install notes.

Quick start

# 1. install
uv venv && uv pip install -e ".[dev]"

# 2. fill your profile
cp profile.example.yaml profile.yaml
$EDITOR profile.yaml        # at least name / target_role
# config.yaml auto-bootstraps from config.example.yaml on first run

# 3. (optional) drop cloud ZIP exports into data/imports/<tool>/

# 4. run pipeline
uv run vibe-resume extract          # parallel extract with progress bar
uv run vibe-resume aggregate        # group by project + infer stack
uv run vibe-resume enrich           # XYZ bullets via claude -p
uv run vibe-resume render -f all    # md + docx + pdf + git snapshot

Commands

Command	What it does
vibe-resume extract [--only NAME]	run extractors, cache to data/cache/*.json
vibe-resume aggregate	group by project, classify task types, infer stack
vibe-resume enrich [-n N] [--locale L] [--tailor JD.txt] [--persona KEY]	generate summary + achievements (XYZ for en, noun-phrase for zh/ja/de/fr/ko); --tailor biases toward a JD's keywords; --persona biases toward a reviewer archetype (see § Reviewer personas)
vibe-resume render -f md|docx|pdf|all [--locale L]	render + git snapshot
vibe-resume render --all-locales [-f FMT]	fan out across every registered locale in one pass
vibe-resume render --tailor data/imports/jd.txt	tailor for a specific job description
vibe-resume review [-v N | --file PATH] [--locale L] [--jd JD.txt] [--persona KEY]	score against the 8-point checklist; --persona appends a persona-specific lens at the end of the report
vibe-resume trend [--locale L]	score history per locale with ASCII sparkline + mean + latest grade
vibe-resume completion {bash|zsh|fish} [--install]	print or install a shell completion script so --locale <tab> expands
vibe-resume status	show per-source activity counts
vibe-resume list-versions / vibe-resume diff 1 2	resume version history

Multi-locale rendering

vibe-resume ships per-locale templates so the same profile.yaml + project data renders into reviewer-appropriate output in each market.

See docs/samples/ for illustrative outputs in en_EU (Europass), ja_JP (職務経歴書), and zh_TW (繁中).

uv run vibe-resume render -f md  --locale en_US     # ATS-optimized US default
uv run vibe-resume render -f md  --locale zh_TW     # 台灣繁中履歷
uv run vibe-resume render -f all --locale ja_JP     # 履歴書 (DOCX grid) + 職務経歴書 (md/pdf)
uv run vibe-resume render -f md  --locale de_DE     # Lebenslauf with Persönliche Daten block

Locale	Style	Photo	Headings	Special
en_US (default)	XYZ action-verb	forbidden	Summary / Skills / Experience / …	Flat ATS-friendly skills line
en_EU	XYZ action-verb	optional	Personal information / Work experience / Education and training / …	Europass-styled — labelled personal-info list, CEFR languages, GDPR-minimal (no DOB by default)
en_GB	XYZ action-verb	forbidden	Personal statement / …	UK spellings, CEFR languages
zh_TW	noun-phrase	optional	自我介紹 / 技能專長 / 工作經歷 / …	全形分隔, 中英技術混排
zh_HK	noun-phrase	optional	Personal Profile 個人簡介 / Work Experience 工作經驗 / …	Bilingual EN + 繁 headings; CEFR; no HKID
zh_CN	noun-phrase	optional	个人简介 / 专业技能 / …	简体, 大厂偏美式
ja_JP	noun-phrase	expected	職務要約 / 職務経歴 / …	DOCX = JIS Z 8303 履歴書 grid (render/japan.py); md = 職務経歴書
ko_KR	noun-phrase	expected	자기소개 / 보유 기술 / 경력 / …	자기소개서 left as separate doc
de_DE	noun-phrase	expected	Persönliche Daten / Berufserfahrung / …	Renders dob / nationality when set
fr_FR	noun-phrase	optional	Profil / Compétences / Expérience / …	1 page jr / 2 pages senior

Per-locale text overrides

UserProfile is extra="allow", so any <field>_<locale> key sits next to its canonical English original and templates pick the right one via the localized Jinja filter:

title: "Senior Full-stack Engineer"
title_zh_TW: "資深全端工程師"
title_ja_JP: "シニアフルスタックエンジニア"

summary: "Full-stack engineer who…"
summary_zh_TW: "全端工程師，熟悉 React / Next.js…"

experience:
  - title: "Senior Full-stack Engineer"
    title_zh_TW: "資深全端工程師"
    company: "Lumen Labs"
    company_zh_TW: "Lumen Labs（種子輪 AI SaaS）"
    bullets:
      - "Reduced query latency from 1.8s to 620ms..."
    bullets_zh_TW:
      - "查詢中位延遲從 1.8 秒降至 620 毫秒…"

Optional locale-conditional personal fields (dob, gender, nationality, mil_service, photo_path, marital_status) are documented in profile.example.yaml. They render only when (a) the active locale's personal_fields list includes them and (b) the value is non-empty.

The full design rationale and per-locale field matrix is in docs/resume_locales.md.

Locale resolution chain

The renderer picks a locale by checking four sources in order and stops at the first one that is set:

1. --locale on the CLI (highest priority)
2. profile.preferred_locale in profile.yaml
3. config.render.locale in config.yaml
4. en_US fallback

# profile.yaml — always render ja_JP unless CLI overrides
preferred_locale: ja_JP

# config.yaml — team default
render:
  locale: en_US
  all_locales_formats: ["md", "docx"]   # what --all-locales produces per locale

Running vibe-resume render --locale zh_TW always wins over preferred_locale; omit it and profile.preferred_locale takes over. The same chain applies when enrich dispatches the LLM prompt, so the right language label is injected regardless of which knob you turn.

Batch rendering every locale

For final-cut day when you need the full pack for different markets:

uv run vibe-resume render --all-locales                 # defaults to config.render.all_locales_formats
uv run vibe-resume render --all-locales -f docx         # force a specific format
uv run vibe-resume render --all-locales --tailor jd.txt # one JD, every locale

--all-locales iterates the full LOCALES registry (currently 10 locales). Per-locale formats are controlled by config.render.all_locales_formats (default ["md"]) — bump it to ["md", "docx", "pdf"] to cut full bundles. --locale and --all-locales are mutually exclusive.

Reviewer personas (--persona)

A résumé that wins a Tech Lead screen often loses an HR-first funnel, and vice versa. --persona adds a third axis (orthogonal to locale and JD) that biases how each bullet is phrased — same candidate, same activity data, but re-voiced for the expected reader.

Key	Reader	What this reader skims for
tech_lead	Staff+ engineer / Tech Lead	Named systems, specific perf numbers, trade-off verbs (migrated / replaced / introduced)
hr	HR manager / recruiter	Career trajectory, collaboration, plain-language business impact; acronym soup gets skipped
executive	VP / hiring manager	Business outcomes in $ / scale / team-size; each role's lead bullet should read as a headline
startup_founder	Early-stage founder	End-to-end ownership, shipping velocity, resourcefulness; enterprise-process framing gets discounted
academic	Research hiring committee	Methodological rigour, datasets, benchmarks, citation-style framing

Compose with locale and JD freely — each dimension hits a different part of the prompt:

# Same person, three audiences, Japanese market
uv run vibe-resume enrich --persona tech_lead   --locale ja_JP -n 3
uv run vibe-resume render  -f all              --locale ja_JP
uv run vibe-resume review  --persona tech_lead --locale ja_JP

uv run vibe-resume enrich --persona hr          --locale ja_JP -n 3
uv run vibe-resume render  -f all              --locale ja_JP     # overwrites the draft
uv run vibe-resume review  --persona hr        --locale ja_JP

# Executive lens + JD-tailored + English
uv run vibe-resume enrich --persona executive --tailor data/imports/jd.txt --locale en_US -n 3
uv run vibe-resume render  -f md --locale en_US --tailor data/imports/jd.txt
uv run vibe-resume review  --persona executive --jd data/imports/jd.txt

Reviewer personas never fabricate — the bias is a re-voicing instruction; numbers, people, and decisions that the raw activity doesn't support stay out. The review --persona form appends a short lens-specific advisory block at the bottom of the scorecard.

Multi-persona prep in one run

Pass comma-separated keys (or all) to produce parallel variants for different audiences without re-extracting:

# Write three persona-specific enrich caches in one command
uv run vibe-resume enrich --persona tech_lead,hr,executive --locale en_US

# Render each persona's variant — filename includes the persona suffix
uv run vibe-resume render --persona tech_lead --locale en_US      # → resume_vNNN_tech_lead.md
uv run vibe-resume render --persona all      --locale en_US      # fan out across every persona

# Side-by-side diff of bullets across personas (quality iteration)
uv run vibe-resume personas-compare -n 3

Each persona writes to data/cache/_project_groups.<persona>.json so variants don't clobber each other. The default _project_groups.json is preserved for persona-less runs. Render filenames include a _<persona> suffix when set.

personas-compare is the quality-iteration tool: it aligns the top-N groups across every persona cache it finds and prints each group's role label + bullets side by side. If two personas produce the same output, the bias isn't biting — revise core/personas.py or the raw activity until they differentiate.

Strategic résumé — target-employer profiles (--company, --level)

70 bundled company profiles and 6 career-level archetypes let you tailor enrich + review output against a specific hiring bar — orthogonal to locale, persona, and JD tailor. Each profile is a YAML at core/profiles/<key>.yaml with fact-check metadata (last_verified_at) so stale data surfaces loudly instead of silently biasing résumés.

Inspect the bundled catalogue:

# Grouped by tier — frontier_ai / ai_unicorn / regional_ai / tw_local /
# us_tier2 / eu / jp / kr
uv run vibe-resume company list
uv run vibe-resume company list --tier jp

# Full profile in human-readable form (must-haves, red flags,
# keyword anchors, enrich_bias, review_tips, verified date)
uv run vibe-resume company show openai

# Age table across all 70 — flags profiles past the 90-day threshold
# (quarterly-refresh cadence matched to current AI-hiring market churn)
uv run vibe-resume company audit
uv run vibe-resume company audit --only-stale --stale-days 90

Apply a company and level when enriching + reviewing:

# Tailor bullets for an OpenAI senior-IC résumé
uv run vibe-resume enrich  --company openai --level senior --locale en_US -n 3
uv run vibe-resume render  -f all --locale en_US
uv run vibe-resume review  --company openai --level senior --locale en_US

# Same activity, different employer — each adds an employer-specific
# keyword-coverage score (0/10) to the review card on top of the 8-point
# base rubric
uv run vibe-resume review  --company anthropic --level senior --locale en_US
uv run vibe-resume review  --company rakuten  --level senior --locale ja_JP

Every --company apply auto-checks verification age. If the profile is older than 90 days, the CLI prints a loud warning and suggests the refresh path — it never silently biases a résumé against stale research.

Refresh a stale profile:

# Delegate a fact-check to the claude CLI agent (max ~5 web queries).
# The report lands in data/verification_reports/<key>_<YYYY-MM-DD>.md.
uv run vibe-resume company verify openai

# If the agent returns VERDICT: clean, auto-bump the verified date:
uv run vibe-resume company verify openai --apply

# After a manual web-browser fact-check, bump the date yourself:
uv run vibe-resume company mark-verified openai
uv run vibe-resume company mark-verified openai --date 2027-01-15 --yes

mark-verified rewrites only the one YAML line, preserving comments, folded-string formatting, and any hand-edited fields. Adding a brand-new profile is drop-in: write core/profiles/<key>.yaml, set last_verified_at: "YYYY-MM-DD", and it is immediately registered, packaged into the wheel, and consumable by --company.

Career levels (--level): new_grad · junior · mid · senior · staff_plus · research_scientist. Each archetype bakes in the lead-bullet signal the reviewer expects at that bracket, so mid-level bullets don't get promoted into staff claims the candidate cannot defend.

Reviewer-view audit (vibe-resume review)

After rendering, score the output against the same 8-point checklist a real reviewer applies:

uv run vibe-resume review                    # latest version
uv run vibe-resume review -v 9               # specific version
uv run vibe-resume review -v 12 --jd jd.txt  # add JD-keyword echo scoring

Each draft is graded on:

1. Top fold — name + target role + at least one concrete metric in the first ~12 lines
2. Numbers per bullet — ≥60% of work-section bullets carry a metric
3. Keyword echo (JD) — JD's top capitalized terms present in the draft (skipped without --jd)
4. Action-verb first — past-tense verb at bullet start (XYZ locales)
5. Density (noun-phrase) — bullets self-sufficient, no dangling pronouns (noun-phrase locales)
6. Red flags — locale-aware photo / DOB / "References available upon request" / consecutive-heading checks
7. Contact line width — header line fits printed page width without ugly wrap (CJK chars count double)
8. Page count — line+wrap-based estimate vs locale target (US/UK ≤2, DE/JP/KR ≤3)

Outputs data/reviews/<draft>_review.md and .json for diffing across iterations. A grade ≥ B/(80%) is the bar before sending the draft to a real reviewer.

Score trend (vibe-resume trend)

Every review run drops a JSON artefact into data/reviews/, and trend folds them into one per-locale summary so you can see whether each market version is improving or regressing as you iterate:

uv run vibe-resume trend               # all locales
uv run vibe-resume trend --locale zh_TW

 Locale  Runs  First    Latest        Mean    Grade  Trend
 en_US   6     58/80    v16: 78/80    91.0%   A      ▂▅▆▇██
 ja_JP   3     50/80    v14: 72/80    82.5%   A      ▁▅█
 zh_TW   4     42/80    v15: 74/80    85.0%   A      ▁▃▆█

The sparkline uses U+2581..U+2588 Unicode blocks so it renders in any monospace terminal. Columns: run count, first score, latest score (with version number), cross-run mean %, grade of the latest run, and the per-run trend.

Claude Code 30-day cleanup — important

Claude Code purges session JSONL transcripts older than 30 days by default. To keep history long-term:

# 1. extend retention
python3 -c "import json,pathlib; p=pathlib.Path.home()/'.claude/settings.json'; \
  d=json.loads(p.read_text()); d['cleanupPeriodDays']=365; \
  p.write_text(json.dumps(d,indent=2,ensure_ascii=False))"

# 2. periodic rsync backup (included)
chmod +x scripts/backup_claude_projects.sh
./scripts/backup_claude_projects.sh
# then register as launchd / cron for weekly backup

Windows backup (Task Scheduler)

scripts/backup_claude_projects.ps1 is the PowerShell-7 equivalent that uses robocopy /MIR /XO to mirror %USERPROFILE%\.claude\projects into %USERPROFILE%\ClaudeCodeArchive\current, with a dated snapshot on the side:

# one-off run
pwsh -NoProfile -File scripts\backup_claude_projects.ps1

# dry-run (also works on macOS/Linux — great for smoke-testing)
pwsh -NoProfile -File scripts\backup_claude_projects.ps1 -WhatIf

# register as a weekly task (Sundays 03:00)
schtasks /Create /TN "vibe-resume backup" /XML scripts\vibe-resume-backup.xml

scripts/vibe-resume-backup.xml is a ready-to-import Task Scheduler template. Edit the <WorkingDirectory> path before importing to point at wherever you cloned this repo. The PowerShell script is linted by PSScriptAnalyzer in CI on windows-latest, and the -WhatIf branch is also exercised there.

Project layout

vibe-resume/
├── profile.example.yaml   # committed template — copy to profile.yaml
├── config.example.yaml    # committed template — auto-copied to config.yaml on first run
├── profile.yaml           # your PII (gitignored)
├── config.yaml            # your extractor paths + privacy rules (gitignored)
├── cli.py                 # entry point (also installed as `vibe-resume`)
├── core/
│   ├── schema.py          # Pydantic v2: Activity, ProjectGroup, UserProfile
│   ├── classifier.py      # 18 task categories with bilingual regex
│   ├── tech_canonical.py  # hard-skill vs domain-tag split
│   ├── stats.py           # rolling window stats (30d/7d)
│   ├── privacy.py         # redaction + blocklist + tech abstraction
│   ├── aggregator.py      # grouping + headline + significance ranking
│   ├── enricher.py        # claude -p → XYZ / noun-phrase bullets per locale
│   ├── review.py          # 8-point scorecard + trend sparkline
│   ├── versioning.py      # git snapshots of drafts
│   └── runner.py          # ThreadPoolExecutor pipeline + rich.progress
├── extractors/
│   ├── local/             # 11 local extractors
│   ├── cloud_export/      # 7 ZIP importers
│   └── api/               # 6 AIGC extractors
├── render/
│   ├── renderer.py        # md / docx / pdf
│   ├── japan.py           # JIS Z 8303 履歴書 grid (ja_JP DOCX path)
│   ├── i18n.py            # LOCALES registry + per-locale label dicts
│   └── templates/resume.<locale>.md.j2
├── scripts/
│   ├── backup_claude_projects.sh       # macOS / Linux rsync
│   ├── backup_claude_projects.ps1      # Windows PowerShell 7 (robocopy)
│   ├── vibe-resume-backup.xml          # Task Scheduler import template
│   └── com.vibe-resume.backup.plist    # macOS launchd agent
├── data/
│   ├── imports/           # put downloaded ZIPs here (gitignored except sample_jd.txt)
│   ├── cache/             # per-source extracted JSON (gitignored)
│   ├── resume_history/    # rendered outputs + internal git (gitignored)
│   └── reviews/           # review reports + history (gitignored)
├── docs/samples/          # illustrative locale-specific sample outputs
├── skills/ai-used-resume/                   # canonical Agent Skill (all 8 hosts, via symlinks)
│   ├── SKILL.md                             # 5-section body, agentskills.io compliant
│   └── references/                          # strategic-resume · troubleshooting · extending
├── .claude/skills/ai-used-resume/           → symlink → skills/ai-used-resume/
├── .gemini/skills/ai-used-resume/           → symlink → skills/ai-used-resume/
├── .agents/skills/ai-used-resume/           → symlink → skills/ai-used-resume/ (Codex + Warp)
├── .opencode/skills/ai-used-resume/         → symlink → skills/ai-used-resume/
├── .claude-plugin/plugin.json               # Claude Code plugin manifest (marketplace)
└── .codex-plugin/plugin.json                # OpenAI Codex plugin manifest (marketplace)

Add a new extractor

# extractors/local/mytool.py
from core.schema import Activity, ActivityType, Source
NAME = "mytool"
def extract(cfg: dict) -> list[Activity]:
    return []  # emit Activity objects

Register in core/runner.py → LOCAL_EXTRACTORS and enable in config.yaml.

Known limits

• Full $HOME scan (git_repos, aider) takes 1-3 min first run even with 4× parallel extractors — switch to scan.mode: whitelist to scope it. _find_repos has a 120s wall-clock deadline to survive FUSE mounts / broken symlinks.
• Grok / Perplexity / Mistral export schemas are lenient parsed (schema not officially published); drop a real sample into data/imports/ if fields don't match.
• Claude Desktop chat contents are encrypted in Local Storage — only MCP config + extensions are extractable.
• PDF rendering requires pandoc + XeLaTeX for CJK; falls back to plain pandoc otherwise.

License

MIT — see LICENSE.

Related projects

Searched the 2026 landscape thoroughly — no direct competitor does all three of (1) multi-source AI-tool extraction, (2) 10-locale rendering, and (3) reviewer audit. The closest adjacent projects cover only one of the three:

• whoisjayd/gitresume — closest adjacent tool; extracts from local/remote GitHub repos → résumé bullets. Single-source (git only), single-language. vibe-resume adds 18 AI-tool extractors, 10 locales, and the reviewer audit.
• AmirhosseinOlyaei/AI-Resume-Builder — generates bullets from PR descriptions and commit messages. Same thesis, narrower source (no Claude Code / Cursor / ChatGPT).
• AndreaCadonna/resumake-mcp — LaTeX résumé via MCP; render-only, no extraction layer. Complementary to our pipeline's render stage.
• javiera-vasquez/claude-code-job-tailor — YAML experience → JD-tailored PDF in Claude Code. Complementary at the JD-match stage; our --tailor is the same idea but starts from extracted activity rather than hand-maintained YAML.

vibe-resume is the only tool in this list that extracts from multiple AI tools + renders in 10 culture-specific locales + includes reviewer audit with trend tracking.