Growth Brain (cog-kb)
Context engineering · knowledge ops
Institutional memory for an AI-assisted GTM team, rebuilt as a git repository — so every Claude session starts from the company's actual strategy, not a blank page.
The team drafts most of its output with Claude — pages, campaigns, decks, case studies, outbound. But every new session started from zero: no memory of the strategy, the approved claims, or which terms were off-limits. That produced drift, unsafe claims, and the same context re-explained every day. Not prompt-engineering problems — knowledge-management problems wearing a prompt-engineering costume.
Built cog-kb: a git-versioned knowledge base treating company knowledge with the discipline of code — single source of truth, a status header on every file, changelog, pull requests, and zero-dependency Python build scripts that run anywhere. Roughly 250,000 words of curated GTM knowledge, plus 15 packaged Claude skills that read from it to produce real deliverables. Governance the system enforces, not a person: an override register that corrects an upstream data source I can't edit, and a governed write-path back into the corpus via pull request.
A new Claude project — or a new hire — boots to full operating context from a single file. Every fact lives in exactly one place, so drift has nowhere to start; brand and claim guardrails apply automatically; recurring work becomes one skill invocation. It's the backbone the rest of the AI-native GTM stack — prospecting, enablement, reporting — runs on.
cog-kb/ ├── CLAUDE.md # Session-start operating contract ├── BOOTSTRAP.md # One-page primer — seeds any new Claude project ├── CHANGELOG.md # ~118 logged entries │ │ ── CONTENT TREE · 76 status-governed files ── ├── meta/ # How to operate the brain ├── company/ # Identity, strategic frame, people ├── market/ # Sector, regulatory landscape, terminology ├── buyers/ # ICP, personas, journey, triggers, objections ├── product/ # Taxonomy, 10 named products, assessment science ├── positioning/ # Narrative, taglines, value props, proof ├── brand/ # Guidelines, voice, language, design system ├── gtm/ # Strategy, pricing, channels, partners ├── evidence/ # Case studies, testimonials, data, validations ├── enterprise/ # Quarantined workplace-platform context ├── tools-and-stack/ # Tools inventory, data pipelines │ │ ── EXECUTION LAYER ── ├── skills/ # 15 packaged Claude skills · 177 files │ │ ── MACHINERY ── ├── Ingredients/ # Transient intake queue ├── resources/ # Curated external source library ├── _build/ # 3 stdlib-only Python build scripts ├── _dist/ # Generated bundles (gitignored) └── _archive/ # Superseded but preserved