— For the Poolside team
You're running real AI on your own laptop now. No jargon here. This is the map of what the pieces are and how they fit, so it stops feeling like magic and starts feeling like a tool.
Read the simple version. Open Under the hood only if you're curious.
01 The mental model
Once these click, everything else is just combinations of them.
The AI, working inside your own files. Not a chat window in a browser. It reads, writes, and builds right where your work already lives.
Claude Code runs in the terminal on your Mac. It can open your project folders, edit code, run commands, and talk to your other tools. The model doing the heavy building is Opus; lighter work runs on Sonnet.
Saved expertise. A skill is a recipe the AI follows the same way every time, so you get a pro result without writing the instructions yourself.
You already have three installed: the humanizer (strips the robotic "AI voice" out of any copy), SEO/AEO (gets you found on Google and in AI answers), and the meeting loop (turns a recording into a clean transcript and a to-do list). Skills live in a folder and sync to every device.
Doorways to your other tools. An MCP lets the AI reach into Canva, GitHub, or your site and actually do the thing, instead of just describing it.
MCP stands for Model Context Protocol. Think of it as a secure plug between the AI and an app. Connect the Canva MCP and the AI can build a design in your Canva account. Connect GitHub and it can push files live. You authorize each one with your own login.
Specialists. When a job needs a focused expert, the AI hands it to an agent that does one thing well, then reports back.
An agent is a skill with a job and a personality. Example: an ADA-compliance agent that only checks a site against accessibility law and flags what would get you sued. You can have a roster of them, each tuned to one task.
The one that powers the rest
A folder that holds everything the AI should know about Poolside — your voice, your clients, your services, the answers to the discovery questions. The AI is only as good as what it knows about you. Fill this once and every task above gets faster and sounds more like you.
It's plain text files in a structured folder. The same idea as Jason's "little brain" vault he showed you. The team feeds it by answering a discovery grill and a values grill, then by recording meetings. It grows on its own as you work.
02 The one distinction worth knowing
Skill
It's knowledge. The "how to do this well" part.
MCP
It's reach. The "go do it in my actual tools" part.
Most real work is a skill and an MCP together: the AI knows how (skill) and can act (MCP).
03 The loop everything runs on
04 Your workflows
What it means: you stop paying the Squarespace-edit tax and stop waiting on a developer.
Today: augment Squarespace with code injection where you already work. Next: a Webflow MCP for true edit-by-prompt, so the team (and your own clients) can add a section without touching code. Static builds host on Cloudflare Pages for free, deployed from GitHub.
What it means: calls stop living in your head. This very page came out of that workflow.
The meeting-loop skill runs Whisper on your Mac (private, no cloud), separates speakers, and writes a transcript plus a two-sided action list. Pair it with WhisperFlow for dictation so you talk instead of type. Put phone calls on speaker and record those too.
What it means: the 4–5 hour Canva job becomes a fill-in. Client reshuffles stop hurting.
Canva Brand Templates plus the Canva MCP for volume and team access; Figma for the higher-craft pieces (it keeps your fonts and layers, and pushes back and forth). Items are treated as data, so a v2 edit is a swap, not a rebuild.
What it means: one pass that lifts traffic and closes the accessibility risk. An easy add-on to resell.
The SEO/AEO skill makes you visible in Google and in AI answers (ChatGPT, Perplexity), and a robots.txt tweak alone moved traffic hard in testing. The ADA agent is being trained on real accessibility law and the lawsuit specifics, so it flags WCAG issues before they cost you.
What it means: a files-in, cuts-out system that hands your editor a strong first cut instead of a blank timeline. We scope this as its own build.
A repository model: footage goes into one place, the system assembles a first cut with your design rules, and an editor finishes from ~70% instead of a blank timeline. Needs a sample of your footage and one finished cut as the style reference.
Ready for your side of the list?
See the action items