// reference · tool stack

The tools, and what to build with them

Here are the real tools teams actually reach for, organized by the job each one does, plus three example stacks so you can see what a build looks like at your size.

The book names no products, on purpose. The models come and go and the architecture is what lasts. Everything on this page will be renamed, acquired, or obsolete before long — that’s fine, because these are temporary means of execution, not the point. Learn the five overlays and the checkpoint discipline from the book; use this page to pick whatever currently walks a leg well. When a product on this list dies, your Map doesn’t.

The six layers of a stack

Every AI build is some combination of these six jobs. You rarely need all of them at once — start with the brain and a human, add layers only as a real workflow demands them. Tap any layer for the named tools and how to pick.

The brain

reasoning model

The model you talk to. ChatGPT (OpenAI), Claude (Anthropic), and Gemini (Google) are the three most teams start with. Microsoft 365 Copilot bundles one into Office; Perplexity is the one for research with sources.

Your own knowledge

grounding / retrieval

Getting the model to work off your documents, not just its training. Start with ChatGPT Projects, Claude Projects, or Google NotebookLM. For bigger corpora, a vector database (Pinecone, Supabase pgvector) with a library like LlamaIndex.

The glue

workflow automation

Connecting the model to your other systems — inbox, CRM, calendar, spreadsheets — so the loop runs without copy-paste. Zapier is the easiest, Make is more powerful, n8n is open-source/self-hostable, Power Automate for Microsoft shops.

Build your own tool

the app layer · own the factory

When renting a tool stops making sense, you build one. No-code: Airtable, Retool, Glide, Bubble, Softr. AI app builders: Lovable, v0, Replit, Bolt. For developers: Next.js + Vercel + Supabase.

Talking & listening

voice, chat, meetings

Capturing and handling conversations. Meeting notes: Otter, Fireflies, Fathom, Granola. Support chat: Intercom Fin, Sierra. Voice agents: Vapi, Bland, Retell. Voice synthesis: ElevenLabs.

The human seat

review & QA · most important

The most important tool in the stack isn't software — it's a named person who checks the output before it ships. Usually the same inbox, doc, or queue the work already lands in. No new tool required; a defined responsibility required.

Three example stacks

What it actually looks like assembled, at three sizes. Steal the one closest to you as a starting point — then cut anything you don't need yet.

The starter stack

Solo founder / small business · roughly $20–100/mo

Brain: ChatGPT Plus or Claude Pro ($20/mo)
Your knowledge: ChatGPT Projects / NotebookLM (included)
Glue: Zapier (free → starter)
Build your own: Airtable or Glide, only if needed
Human seat: You, in your own inbox

You can run your first handful of automations on this for the price of one team lunch a month. Don't add anything until a real bottleneck demands it.

The growing-team stack

10–100 people, a few functions automating

Brain: ChatGPT Team or Claude Team (per seat)
Your knowledge: Projects + a shared knowledge base; NotebookLM
Glue: Make or n8n for multi-step workflows
Build your own: Retool / Bubble for internal tools
Talking & listening: Fireflies or Fathom for meetings
Human seat: A named reviewer per function

This is where the 12-Week Program lives: one builder per function, patterns transferring across teams, review owned by a named person. Mostly off-the-shelf, no engineering required.

The enterprise / regulated stack

Governed, in your own cloud, audit-ready

Brain: Azure OpenAI, Claude on Bedrock, or Vertex AI
Your knowledge: Vector DB (Pinecone / pgvector) + LlamaIndex
Glue: Power Automate or self-hosted n8n
Build your own: Internal dev: Next.js + Vercel + Supabase, plus Retool
Governance: SSO, audit logging, human-in-the-loop gates
Human seat: Compliance + a named owner per workflow

The models run inside your own cloud tenancy so data and audit trails stay under your control. Pair this with the Compliance & Legal playbook — what you can't afford to get wrong never runs without a checkpoint.

Tool names and prices are real examples as of this writing, not endorsements or rankings — and they change fast. Pick by the job and your situation, run it on real work, and switch only when a specific gap appears. The layers don't change; the logos in them will.

From Running at the Speed of Light, On tools — a deliberate omission · Prometheus Press