seed-workspace

A thinking machine at your fingertips. Connect it to your systems — meetings, CRM, email, calendar, drive — and it handles the retrieval and processing so you don't have to. You get a universal interface across everything you work with, available the moment you open a chat.

The AI's job is to hold the context, do the synthesis, and execute. Your job is the part humans are actually good at: wisdom, taste, direction, vision. The seed is structured to keep that division clean.

The best version found so far of having all the knowledge, context, and tools in one place. Not a productized template. Not a wizard. A methodology + a folder structure with intent + a few skills + an optional substrate. You shape the rest.

Dynamic, file-versioned, cross-system — the alternative to Claude Projects / Custom GPTs / Gems, which are sealed environments with no cross-system access, no git history, and no extension path. This is the opposite of all three.

What you can do in a session

This workspace is designed for human-in-the-loop leverage — not background automation. You direct; the AI synthesizes across systems at a speed no human-only workflow matches.

Tasks that used to take an hour:

• Pre-call prep: pull the full conversation history with a person or company, brief yourself, surface open threads — before you join
• Project context update: take recent transcripts, extract action items and discussion points, drop them into a spec or context file
• CRM update: log a call or email chain with full notes — in seconds, with everything visible for review before it's written
• Email with full context: draft a reply while pulling the entire relationship history into scope
• Sales coaching: review a series of conversations for patterns, objections, and moments worth studying
• Cross-system report: pull from transcripts + CRM + calendar + drive, synthesize, save as a working file

Higher trust and quality of output with you in the loop — because you direct the synthesis, catch the edge cases, and make the judgment calls the AI can't.

The deeper value: by offloading retrieval and processing, the seed frees your working memory for the things that actually require a human — reading the room, making the call, knowing what matters. Deep thinking becomes possible when you're not also doing information logistics.

Work isn't ephemeral. The failure mode of inline chat is that the best thinking lives in a conversation window and evaporates. Here, useful outputs get captured as files, promoted to project contexts, or integrated into canonical docs via an explicit approval step. The AI's work product persists.

Any task can become a skill. Any project can carry its own AGENTS.md. When you find yourself describing the same workflow twice, that's a skill waiting to be written. The extension mechanism is structural — same architecture as the seed itself.

---

Read in order:

1. PHILOSOPHY.md — what this is and why
2. WORKFLOWS.md — the four feedback loops
3. OnboardingChecklist.md — one-screen adoption guide
4. RECOMMENDED-TOOLING.md — short menu of things that pair well (includes OneCLI for API key protection)

Security baked in: real API keys do not belong in files the AI reads (mcp.json, committed config, chat). Store them in a credential gateway (recommended: OneCLI); agents use placeholders or proxy-routed access. See PHILOSOPHY.md (Credential security) and AGENTS.md.template.

---

What's here

seed-workspace/
├── PHILOSOPHY.md               ← what this is and why
├── WORKFLOWS.md                ← the four feedback loops the seed runs on
├── RECOMMENDED-TOOLING.md      ← short menu, not a prescription
├── OnboardingChecklist.md      ← one-screen adoption guide
├── Setup-Decisions.md          ← per-ability decision matrix (output of discover-stack)
│
├── Context/                    ← stable layer: self-knowledge, positioning, voice
├── Strategy/                   ← working layer: synthesis + Evidence Ledger + staging
│   ├── to integrate/           ← capture-then-integrate staging area
│   └── z-archived/             ← superseded artifacts kept for trail
├── Container/                  ← operational layer: weekly plans + cadences
├── In-App/                     ← system-prompt template for tools without auto-loaded context
│
├── clients/                    ← symlinks to per-engagement repos (sibling on disk)
├── personal/                   ← symlinks to non-client personal projects
│
├── infra/
│   ├── security/               ← gitleaks + OneCLI setup for this workspace
│   ├── transcript-vault-mcp/   ← optional substrate: local meeting KB, recorder-agnostic
│   └── drive-symlinks/         ← convention for surfacing Drive content as local files
│
├── scripts/
│   ├── secret-scan.sh          ← local gitleaks (full history)
│   └── install-git-hooks.sh    ← optional pre-commit secret scan
├── .gitleaks.toml              ← gitleaks config (CI + local)
├── .github/workflows/          ← secret-scan.yml on push/PR
├── .cursor/mcp.json.template   ← MCP wiring template
└── .claude/skills/             ← the seed's skills

The technical substrate: transcript-vault-mcp

infra/transcript-vault-mcp/ is the most substantial piece of working infrastructure in the repo. It's a local-first semantic search MCP built from your meeting recorder's data — no third-party AI API required for retrieval, everything runs on disk.

What it does that your recorder's UI can't:

• Semantic search across every transcript you've ever recorded, in one query
• Pull the full history of every conversation with a person or company, chronologically ordered
• Verify a claim by retrieving both supporting and contradicting evidence (dual-retrieval)
• Infer and track project/deal status from mentions across meetings over time
• Build pre-meeting briefings, client personas, and sales coaching reviews as MCP workflow prompts

Stack: TypeScript, SQLite + FTS5 (keyword search), LanceDB + @xenova/transformers (local 384-dim embeddings — no OpenAI dependency), MCP stdio server, adapter interface for any recorder. Fathom is fully implemented; Fireflies/Otter/Granola ship as stubs.

Originally built for Fathom specifically as opsMachine/FathomMCP, then generalized so the extraction layer is swappable while everything else (transform, embed, search, MCP server, workflow prompts) stays recorder-agnostic.

→ Full documentation: infra/transcript-vault-mcp/README.md
→ Architecture deep-dive: infra/transcript-vault-mcp/ARCHITECTURE.md

---

The skills (the methodology, codified)

• harvest-learnings + integrate — capture-then-integrate lifecycle with explicit per-edit approval gates
• gain-analysis — the central build/skip decision protocol
• discover-stack — day-1 interview to map what you already have so we don't recreate
• scaffold-cadence — install one operational cadence at a time

Skill paths are parameterized via .claude/skills/.skill-config.yml so the seed adapts to whatever you've named your canonical docs.

Getting started

1. Clone (or "Use this template" on GitHub) to ~/Documents/GitHub/<your-meta-workspace-name>/.
2. Open it in your IDE (Cursor, AntiGravity, Claude Code) as the workspace root.
3. In chat: "discover my stack" — invokes discover-stack, which interviews you and writes Setup-Decisions.md. It will ask, for each canonical doc, whether you already have it somewhere (Drive, Notion) so we symlink rather than recreate wherever possible.
4. From there, your AI walks you through OnboardingChecklist.md at your pace.

A few things the AI may need to do for you (just ask): create the .cursor/skills → ../.claude/skills symlink for IDE skill discovery; create empty vault/People/ and vault/Companies/ if you decide to enable transcript-vault-mcp; create symlinks under clients/ and personal/ as you add sibling repos. None of these need a script — they're one-line shell commands the AI can run on demand.

What this is NOT

• Not a wizard. Onboarding is a one-screen checklist; nothing blocks.
• Not opinionated about your stack. Every "should I build / connect this?" routes through gain-analysis.
• Not a clone of someone else's setup. Templates are starting points; the default lean is symlink to what you already have rather than fill in blank docs.
• Not finished. v0.1. The whole point is infinitely extensible — adapt freely, replace anything, write your own skills, build your own MCPs.

IDE notes

Default .cursor/ folder works for Cursor + AntiGravity. For Windsurf, Claude Code, or others, see .cursor/README.md — it's usually one symlink.

Recommended plugins (all available in this workspace): Excalidraw editor, PDF Previewer, Markdown Preview Enhanced, and Word Document Viewer — see RECOMMENDED-TOOLING.md for extension IDs. The AI can also read, create, and edit .excalidraw files directly; the excalidraw-pyramid-parser skill handles iterative strategy diagram refinement.

License

MIT — see LICENSE.