Company Second Brain
jmchalayer/company-second-brain
Base de conhecimento centralizada e estruturada para times trabalhar com IA
Ver no GitHubUm starter para construir a camada de conhecimento compartilhado da sua equipe, estruturado em três níveis (mapa, wiki e dados). Funciona como repositório centralizado conectável a qualquer ferramenta de IA que o time use (Claude, ChatGPT, Cursor), eliminando a necessidade de buscar contexto em múltiplos lugares.
Instalação
npx convez-ai jmchalayer/company-second-brain1 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
- Criar repositório único de contexto para agentes Claude e outras ferramentas de IA
- Estruturar conhecimento de marca, produto e clientes de forma IA-readable
- Onboarding de novos membros da equipe com acesso imediato ao contexto compartilhado
- Processar transcripts de calls e meeting para insights estruturados
Veredito

Solução bem pensada para times que usam IA regularmente. O design em três camadas (mapa → wiki → dados) é eficiente e evita sobrecarga de contexto. Exige disciplina na manutenção, mas paga dividendos em velocidade de trabalho com agentes.
README
A tradução para o português deste README ainda está na fila.
Company Second Brain
A starter for building your company's shared knowledge layer. The companion to founder-second-brain - that one is your personal AI accountability partner. This one is your team's.
Why a company brain
Your personal brain works for you. But once a team forms, the real bottleneck is everyone else: customer support, content, ops, sales, founders' calls. They all need the same context (who you are, what the brand sounds like, what your customers say, how you talk about pricing) and most of it lives in your head, in Slack, or in scattered Notion docs.
A company brain fixes that. One folder structure. One source of truth. Connectable from any AI tool the team uses.
How it works
Claude Code, Codex, Cursor, ChatGPT,
custom GPTs, anything next
│
▼
┌──────────────────────────────┐
│ CLAUDE.md / AGENTS.md │ ← the map
├──────────────────────────────┤
│ Wiki/ │ ← synthesis layer
├──────────────────────────────┤
│ Brand, Company, Customers, │ ← data layer
│ Content, Product, Strategy, │
│ Support, Growth, People │
└──────────────────────────────┘
Three layers, each with a job:
- Map (
CLAUDE.md/AGENTS.md): tells the agent what's in the vault and where to look. Read first, every time. Tiny. - Wiki: a synthesised, cleaned-up summary of each domain. The agent reads this when it needs an overview.
- Data: the raw stuff. Folders by department. The agent only opens specific files when it has a specific reason.
This is what makes it fast. The agent doesn't load the whole vault. It reads the map, decides what to look at, and pulls only what it needs.
What's in the box
company-second-brain/
├── README.md # this file
├── CLAUDE.md # the map (read first)
├── AGENTS.md # mirror, for non-Claude agents
├── Brand/ # voice, design system, brand bible
├── Company/ # mission, values, product overview, team
├── Customers/ # ICP, segments, interview transcripts
├── Content/ # voice guide, philosophy, briefs, outputs
├── Growth/ # campaigns, channels, attribution
├── People/ # team, advisors, key contacts
├── Playbooks/ # how-we-do-X documentation
├── Product/ # roadmap, decisions, features
├── Strategy/ # pricing, positioning, plans
├── Support/ # FAQ, common scenarios, tone
├── Templates/ # templates referenced by skills
├── Transcripts/ # raw call/meeting transcripts (pre-processing)
├── Wiki/ # agent-readable synthesis layer
├── Logs/ # session summaries
└── .claude/skills/ # 4 starter skills (see below)
The 7 starter skills
| Skill | What it does |
|---|---|
/onboard | First-time setup. Reads your website, blog, pricing page, any docs you upload. Pre-fills brand voice, design system, ICP, product overview, pricing - section by section, you confirm or correct. About 30 minutes start to finish. |
/transcript-processing | Process customer call or team meeting transcripts. Extracts insights, updates relevant Wiki pages, files raw transcripts in Transcripts/. |
/update-wiki | Refresh the Wiki layer. Run after big changes. Reads the data layer, regenerates synthesis pages. |
/prd | Write a product requirements doc using a 6-section framework. Gathers the brief conversationally, drafts, saves to Product/prds/. |
/humanizer | Strip AI-isms from drafts before publishing. Detects 29 patterns (filler phrases, em dash overuse, promotional language, rule of three, etc.). |
/link-sweep | Vault hygiene. Audits red wikilinks, broken frontmatter, orphan files. Fixes the mechanical, flags the judgement calls. |
/write-skill | Meta. Add your own skills (CS replies, landing-page builder, content drafter, whatever your team needs). |
Requirements
- Any LLM tool that reads files and runs slash commands. Tested with Claude Code, Codex CLI, and Cursor. Should work in anything that respects
CLAUDE.mdorAGENTS.md. - Optional: web search / fetch tools for the
/onboardskill (Claude Code has these by default).
Install
git clone https://github.com/jmchalayer/company-second-brain.git our-vault
cd our-vault
Then in your AI tool of choice:
- Open the folder.
- Run
/onboard. - Drop your company URL when asked. Optionally paste a brand bible, design tokens JSON, or Figma exports if you have them.
- Confirm or correct each section.
- Share the folder with your team (Drive, Dropbox, GitHub, your call).
How the team uses it
After setup, anyone on the team can:
- Ask any LLM "draft a CS reply to this customer" - it'll use the Support tone guide and FAQ automatically.
- Ask "draft a LinkedIn post about our AI launch" - it'll use the Brand voice and Content philosophy.
- Process a customer call: drop the transcript, run
/transcript-processing, the insights flow into Customers/. - Add a new skill: run
/write-skill, codify a workflow your team does often.
What's deliberately not here
- Specific integrations (Linear, Notion, HubSpot, Stripe). Add them yourself. The vault holds pointers to those systems, not duplicates.
- Auth, permissions, granular access control. That's your file-sharing system's job (Drive, GitHub, Dropbox). Treat the whole vault as team-readable.
- Per-person memory. Use the personal founder-second-brain for that.
Customising
Everything in .claude/skills/*/SKILL.md is plain markdown. Edit the questions, change the tone, add your own steps. The skills are opinionated by default and easy to override.
Companion repo
Your personal brain lives separately. Get the founder version: github.com/jmchalayer/founder-second-brain.
Licence
MIT. Do whatever you want with it.
Credits
Inspired by Andrej Karpathy's LLM Wiki pattern. The architecture (map → synthesis → data) and the skill design come from running this in production at a real company for six months.
Built by JM Chalayer, founder of JENA.
Want help adapting this to your own workflow, team, or training programme? Find me on LinkedIn or Instagram.