Career Database
pgrav001/career-database
Base durável para sua carreira: claim → proof em markdown
Ver no GitHubUma skill Claude Code que constrói uma base de dados durável de carreira usando o modelo Claim → Proof. Cria uma estrutura de markdown que funciona como substrato para gerar resumes, conteúdo LinkedIn, cartas de apresentação e preparação de entrevistas, mantendo tudo organizado e reutilizável.
Instalação
npx convez-ai pgrav001/career-database1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Consolidar histórico de conquistas e evidências para múltiplos ciclos de busca de emprego
- Gerar resumes, LinkedIn About e cover letters a partir de um único source of truth
- Preparar respostas estruturadas para entrevistas com narrativa consistente
- Manter registro operacional de oportunidades, decisões e questões em aberto
Veredito

Solução bem arquitetada para quem busca profissionalismo em gestão de carreira. O modelo Claim → Proof é sólido e o sistema de templates + convenções reduz fricção entre sessões. Recomendado para quem trabalha regularmente com candidaturas e quer deixar a aleatoriedade de lado.
README
A tradução para o português deste README ainda está na fila.
career-database
A Claude Code skill for building a durable, AI-native career database — the substrate that resumes, LinkedIn copy, cover letters, and interview prep are generated from, not maintained alongside.
Built around a Claim → Proof narrative model. Strengths are the headline claims. Evidence and stories are the proof. Peer quotes are the third-party proof. Voice files capture tone. Chronology supports the claims but doesn't lead them.
Why this exists
Most job-search workflows do one of two things:
- Bounce between ad-hoc resume drafts in Google Docs / Word, losing context every cycle, or
- Bolt onto a single tool (LinkedIn, Resume.io, Notion template) that owns the format and constrains the narrative.
Neither builds a durable substrate. The next role, the next promotion conversation, the next interview cycle starts over.
This skill bootstraps a markdown directory that is the substrate. Resumes, LinkedIn About sections, cover letters, and interview answers are generated from it. The substrate gets richer every session; the artifacts are downstream.
The architecture is opinionated. The conventions are explicit. The hygiene rules are codified so future-you (or any AI assistant) can pick up where you left off without re-asking what's already known.
What you get
- A markdown directory structure with
~25 file templates, frontmatter schemas, and conventions - Operational files (
SESSION.md,OPEN_QUESTIONS.md,DECISIONS.md,OPPORTUNITIES.md) that survive across sessions - A Tier 1–5 prioritization framework for "what should I work on first"
- A
Claim → Proofmodel that distinguishes how you're effective from the chronological proof - Hygiene rules (provenance, as-of dating, Open Recall promotion) that prevent drift
- The skill instructs Claude on the read-order, session ritual, and generation flow
Install
career-database is a Claude Code plugin. After cloning the repo to your plugins directory, Claude Code picks it up on next launch and exposes it as the /career-database slash command.
Quickest path — clone directly into the plugin directory:
# macOS / Linux
git clone https://github.com/<your-username>/career-database ~/.claude/plugins/career-database
# Windows (PowerShell)
git clone https://github.com/<your-username>/career-database "$env:USERPROFILE\.claude\plugins\career-database"
The plugin should resolve to:
~/.claude/plugins/career-database/
.claude-plugin/
plugin.json
README.md
SKILL.md
docs/
ARCHITECTURE.md
BOOTSTRAPPING.md
CONVENTIONS.md
WORKFLOWS.md
template/
(~25 markdown files mirroring the database structure)
After installing, restart Claude Code so the plugin registry picks up the new skill.
If you don't have Claude Code, the templates and docs are still useful standalone — clone the repo and copy template/ to your chosen location. The methodology (Claim → Proof, Tier 1–5, hygiene rules) works in any tool that reads markdown.
Quick start
Once installed and Claude Code has been restarted, invoke the skill in any Claude Code session:
/career-database
The skill will:
- Ask where to put the database (default
~/career/) - Copy template files there if the directory doesn't exist; read the existing operational files (
SESSION.md,OPEN_QUESTIONS.md) if it does - Walk you through the first session — Tier 1 interview-blocking questions first (see
docs/BOOTSTRAPPING.mdfor the full first-session guide)
Try it before you commit
If 90–120 minutes of bootstrap feels like a lot, two lower-cost ways to evaluate first:
- Read the worked example at
examples/sample-arc/— a fully-fictional populated database (15-year arc of a fictional library-system director) showing what evidence files, stories, peer quotes, and operational rhythm look like in steady state. ~15 minutes to read. - The 30-minute MVP path in
docs/BOOTSTRAPPING.md— produces one generated LinkedIn paragraph from one evidence file and one peer-quotes file. Skips most of the ritual; tells you fast whether the methodology produces output worth scaling.
Architecture overview
See docs/ARCHITECTURE.md for the full Claim → Proof model. Short version:
| Layer | Files | Role |
|---|---|---|
| Claims (how you operate) | strengths.md, themes/*.md | Headline material for resumes / LinkedIn / cover letters |
| Proof (what you did) | evidence/*.md | Quantified accomplishments backing the claims |
| Proof in narrative form | stories/*.md | STAR-format behavioral interview material |
| Third-party proof | voice/peer-quotes.md | Curated quotes from managers, peers, reports |
| Voice | voice/self-voice.md | Your own writing patterns for tone-matching |
| Chronology | history/*.md, career-arc.md | Background context, not headline material |
| Operational | SESSION.md, OPEN_QUESTIONS.md, DECISIONS.md, CLAUDE.md | Session continuity + hygiene |
Conventions
See docs/CONVENTIONS.md for full detail. Highlights:
- Frontmatter on
history/,evidence/,stories/files withstatus,confidence,metric, etc. [[wiki-links]]to cross-reference files (human/AI navigation aid; no tooling required)- Provenance flags on every claim:
confirmed/approximate/claimed - As-of dating on every time-sensitive number (
direct-report sentiment 95% (H2 2024), notdirect-report sentiment 95%) - Open Recall sections per file for in-context half-thoughts; cross-cutting questions promote up to top-level
OPEN_QUESTIONS.md
Workflows
See docs/WORKFLOWS.md for full detail. Major ones:
- Source mining: drop raw artifact (resume, performance review, LinkedIn export) into
sources/inputs/, add tosources/inventory.md, mine into the right layers - Interview / brain-dump: when the user wants to brain-dump an era or project in real time — the AI is scribe + interviewer, capturing verbatim into voice files and flushing at natural breakpoints
- Artifact generation: lead with claim → pull proof → match voice → reference theme framing
- Audit / consistency check: periodically compare the database against source materials to surface drift, missing items, framing inconsistencies; report-only deliverable
- Session ritual: start by reading
SESSION.md+OPEN_QUESTIONS.md; end by updatingSESSION.md+ moving answered items + logging structural decisions inDECISIONS.md
Bootstrapping
See docs/BOOTSTRAPPING.md for the first-session guide.
First session priorities:
- Tier 1: interview-blocking questions — answer these before generating any outward-facing artifact. Why did you leave your last role / why are you looking now / what are you looking for / comp expectations / geography.
- Source mining — process whatever you've got (current resume, performance review history, LinkedIn export).
- Substance-layer brain-dump — era-by-era walk through your career; capture evidence + stories + voice as you go.
Privacy & naming other people in your database
A career database is, by design, a record of your career — but careers don't happen alone. Your evidence files will name managers, direct reports, peers, and external partners. Your peer-quotes file will hold things people said about you in private channels (Slack DMs, performance reviews, emails). Your stories will reference people by name when the narrative requires it.
Some of that content is fine. A lot of it isn't fine to publish without thought.
Two layers to keep distinct:
-
The database itself is local. It lives in your filesystem. Treat it like a private journal: don't sync it to anywhere you don't fully control; don't put it in a shared cloud folder; don't commit it to a public git repo. If you back it up, encrypt the backup. (The skill instructions do not send your content anywhere by default — but anything you ask an AI assistant to do with the content involves the assistant reading it. Anthropic's data policies apply to Claude sessions; check them if uncertain.)
-
Anything generated from the database (resumes, LinkedIn copy, cover letters) is public-facing. The substrate is private; the artifacts are not. Before any artifact ships:
- Quoting someone: get their explicit consent, or use a generic attribution ("a senior leader I worked with" rather than "Pat Chen"), or paraphrase rather than quote verbatim. The
voice/peer-quotes.mdtemplate has fields for "OK to quote in artifacts" — use them. - Naming someone: even without a direct quote, naming a person in your career narrative attaches them to your story. Default to roles/relationships ("my system director," "a direct report") rather than names, unless naming them is genuinely load-bearing and they've signed off.
- Internal metrics, dollar figures, dates: judge what's defensible to publish. Some metrics are already public (you mentioned them on LinkedIn; they're in a press release; they're in a published annual report). Others are internal-only. Resumes that recite specific internal numbers without considering this can create real downstream issues, including with previous employers.
- Quoting someone: get their explicit consent, or use a generic attribution ("a senior leader I worked with" rather than "Pat Chen"), or paraphrase rather than quote verbatim. The
The worked example at examples/sample-arc/ demonstrates this layer in action — voice/peer-quotes.md has explicit "OK to quote externally?" notes per quote, the evidence files cite source-grade vs. subject-attested provenance, and the Open Recall sections flag specific items that need verbatim verification before any external use. Use it as the pattern.
What's NOT in here
- Any personal content. Names, companies, metrics, dates, quotes — all yours to fill in.
- A web app, a database, a backing service. This is markdown files. Use any editor, any tool, any AI assistant.
- A specific resume format or output template. The substrate is durable; the artifact templates live elsewhere.
Related
career-database is the substrate layer — durable, broad, owned by the user. A separate skill called autowrite covers the artifact-generation side: given a resume and a list of target companies, it autonomously scores the resume against company-specific binary hiring evals, mutates one line at a time, and produces opening-tailored variants. The two skills are independent (no code-level dependencies) but compose cleanly — career-database builds the substrate; autowrite operates on the canonical artifact produced from it.
If you want the substrate layer (interview prep, voice capture, durable claim → proof structure across your career): you're in the right place.
If you want autonomous resume tailoring across multiple target companies: that's autowrite. Either or both, in either order.
License
MIT. See LICENSE.
Contributing
Issues and pull requests welcome. Especially:
- Additional theme patterns (the strengths-deep-dive layer is the most opinionated and benefits from more examples)
- Frontmatter schema refinements
- Workflow patterns for specific industries / disciplines
- Translations / non-English templates