The Promise
Wire all four surfaces and your AI stack stops being a chatbot you visit and starts being an operating system that runs in the background of your business.
The One-Sentence Setup
Most operators think their AI setup is weak because they're using the wrong model — it's actually weak because they only have two of the four surfaces wired together.
The Core Insight
Every working AI ops setup has exactly four surfaces: Capture (where input lands), Reason (where thinking happens), Act (where AI does things in the world), and Remember (where everything persists). The compound value isn't in any single surface — it's in the wiring between them. A standalone Claude window is one surface. A voice memo that gets transcribed, analyzed, turned into three calendar invites, and filed into a searchable vault is four. The math is multiplicative, not additive. Two surfaces gives you a tool. Three surfaces gives you a workflow. Four surfaces gives you a second brain that operates while you sleep.
The Mechanism
1. Capture — make the input frictionless
What: Every input pipe into your AI stack — voice, text, photo, calendar event, email, browser highlight, Slack message. How: Pick one default capture surface per input type and stop fragmenting. Voice goes to one app. Web reading goes to one reader. Inbox triage lives in one client. The capture surface should have lower friction than the friction of not capturing — under three seconds, no decisions, no folder picking. Miss this and: the operator spends the day thinking thoughts that never enter the system, which means Reason and Remember have nothing to work with.
2. Reason — make the thinking compoundable
What: The model layer that actually processes input — Claude, GPT-4, a local model, or the tool wrapping one of them. How: Pick one daily-driver model and one heavyweight. We use Claude Sonnet 4.6 for the 80% case and Opus 4.7 for deep multi-hour sessions. Stop model-shopping for every task — context-switching cost is higher than the marginal quality gain. Wire Reason directly to Capture so input flows in without copy-paste. Miss this and: the operator becomes a model-comparison hobbyist instead of an operator with a working system.
3. Act — make the output leave the conversation
What: The surface where AI does something in the world — sends the email, posts the update, books the room, runs the query, deploys the code, files the receipt. How: Start with one trusted Act path and harden it. A Zapier zap, a Claude Code bash session, a Raycast script, an Apple Shortcut. The bar is: the operator approved the spec, AI executed it without a second prompt. Most stacks die here because Act is where the trust gate sits. Miss this and: every AI output ends in copy-paste, which means the operator is the bottleneck on every task the AI just "completed."
4. Remember — make the past compound
What: The persistent store where Capture, Reason output, and Act receipts get written. Obsidian vault, Notion database, Supabase pgvector, the file system. How: One canonical store with structured folders and a write-by-default policy. Every voice memo, every Claude session of consequence, every Act receipt gets a file. Searchable. Linkable. Reference-able by the next Reason call. This is the surface 90% of operators skip — and it's the one that separates a tool from a system. Miss this and: every conversation starts from zero, the operator re-explains context daily, and the AI never gets smarter about the specific business.
The Pitfalls
The Two-Surface Trap. Capture + Reason only — "I talk to ChatGPT." Fix: add one Act path this week, even a single Zap, so something leaves the chat window.
The Tool Hoarder. Eleven Capture apps, four models, no Remember. Fix: pick one tool per surface and delete the rest from the dock for thirty days.
The Remember-Last Mistake. Building Capture → Reason → Act first and promising to "set up the vault later." Fix: stand up Remember on day one, even if it's a single dated markdown file per session. Future-you can't index conversations that were never saved.
The Trust Skip. Wiring Act before the operator has watched Reason work for two weeks. Fix: run the full flow with a human-in-the-loop approval step until the false-positive rate is near zero, then drop the gate.
The Surface Sprawl. Three Capture apps, two models, four Act paths, two vaults. Fix: every surface gets exactly one default. Alternates are allowed only when the primary genuinely can't do the job.
The Drill (this week)
Open a blank document. Draw four columns: Capture, Reason, Act, Remember. Under each, list every tool currently in active use for that surface. Time-box this to ten minutes — if a tool isn't top-of-mind, it's not active.
Now score honestly:
- 0 tools in a column: that surface is dead.
- 1 tool, used daily: that surface is wired.
- 2+ tools, none used daily: that surface is fragmented.
Most operators discover one dead surface (almost always Remember) and one fragmented surface (usually Capture). The drill's whole purpose is to surface that. The next move is one of three: stand up the dead surface with the cheapest possible v1, collapse the fragmented surface to a single default, or wire two existing surfaces that aren't talking yet. One move. This week. Not all three.
The Tools
| Surface | Primary | Alternates |
|---|---|---|
| Capture | iPhone Voice Memos | Otter, MacWhisper, Superhuman, Readwise, browser highlight extensions |
| Reason | Claude desktop (Sonnet 4.6 / Opus 4.7) | ChatGPT desktop, Claude Code, Cursor, AI Gateway for routing |
| Act | Claude Code with bash | Zapier, Make, Vercel functions, Apple Shortcuts, Raycast scripts |
| Remember | Obsidian vault | Notion, Supabase pgvector, Mem, structured local file system |
Verala stays vendor-neutral on most of this. The opinion: pick one per row, run it for thirty days, switch only if the surface is genuinely failing — not because something newer launched.
Cross-references
The 4-Surface AI Stack is the umbrella. Every other framework in the Operator Stack plugs into one of these four surfaces:
- The Connector Stack powers Act + Capture — the MCPs and integrations that let AI reach into other systems.
- The Operator Vault is the canonical Remember implementation — an Obsidian setup tuned for founders running real businesses.
- The Voice Memo → AI Loop is the canonical Capture → Reason → Act flow — the end-to-end pipeline that turns a thirty-second walk into three filed actions.
- The MCP Mental Model and The Memory Architecture are the underlying primitives that make Act and Remember actually work.
Read this framework first. The rest are implementations of one of its four columns.
One framework. One drill. One week at a time.
The Operator Stack is the architecture. Verala is the practice that runs it on your own communication delivery — voice, pitch, pause, presence. One foundation per week, until it's automatic.
Take the free 5-Foundation Voice Audit → · Book a 30-min intro call →