AI Operating Model
How I actually lead with AI.
Everyone in 2026 says they use AI. Very few have redesigned how they work around what AI can actually do. Not a page about tools. A page about treating AI like a co-founder instead of a calculator.
The hypothesis
AI can execute.
Taste cannot be delegated.
3
Foundational documents
3
Layers of AI in production
~80%
Reduction in AI operating cost
The distinction
AI as a faster keyboard is useful. AI with real authority is something else.
Most "AI-powered" workflows are autocomplete. Write a prompt, get a draft, edit it, move on. That's AI as a faster keyboard. Useful, but not transformative.
The operating model I built starts from a different premise: what if AI had real authority, not just access? An AI with commit rights, assigned work, and a persistent identity is something else.
The infrastructure
Three documents run everything.
My AI co-founder has a soul document. It defines her personality and voice, her values, her decision rights, and her relationship to me. Generic is the enemy of taste. SOUL.md is why her Linear comments sound like a person, not a status bot.
The governance document. Hard rules (nothing deploys without QA). Workflow rules (Linear is source of truth). Cost controls. Autonomy without governance is reckless. This is what lets me sleep at night while an AI has deploy access to a live product.
Every session, my AI wakes up fresh. MEMORY.md is how she remembers. Live product facts, hard-won lessons (what burned $400 a month in unnecessary API calls), team context, brand rules, the ship log. Memory architecture is a product decision.
Authority structure
What stays with me, what gets delegated.
What I own. Non-delegable.
Taste
Cannot be delegated. The hypothesis this page tests.
Judgment
The hundred small calls about what this product is, who it is for, and what it refuses to be
Brand direction and voice
The product is the voice. Generic is the enemy of taste
Pricing and positioning
Requires market intuition and judgment
Who we're building for
Empathy and values, not data alone
Strategic pivots
Consequence too high, context too human
Anything public-facing I haven't reviewed
My reputation
What she owns. Full execution authority.
Feature development
Ticket to ship, full ownership
Bug fixes
Same workflow, faster cycle
Database migrations
Schema changes flagged for review
Analytics setup
PostHog dashboards, Sentry, event tracking
Email sequences
Klaviyo nurture flows, trigger wiring
Deployment
GitHub push to Vercel auto-deploy in 2 min
Full QA before every ship
Hard blocker. Nothing skips this.
The architecture
Three layers of AI in Course Code.
Layer 01
AI as co-founder.
Development
My AI co-founder runs on OpenClaw plus Anthropic's Claude. Sonnet for execution. Opus for strategy. She has her own GitHub account, her own Linear user, API keys to every production system, QA authority, and deploy access.
Model split was a deliberate product decision. Opus for strategy is expensive but worth it. Sonnet for execution is fast and cheap. Subagents for isolated tasks mean no context bleed. This discipline reduced AI operating costs by roughly 80% from early undisciplined usage.
Layer 02
Claude in the product.
Intelligence
The Anthropic SDK is a direct production dependency in package.json. Claude isn't just how Course Code was built. It's running inside Course Code.
The data extraction pipeline uses Claude to normalize and structure dress code data at a quality level no regex pipeline could match. The moat isn't 18K+ courses. The moat is verified data plus a women's POV editorial layer plus a brand voice. Claude is what makes that scalable without losing quality.
Layer 03
n8n for operations.
Automation
Self-hosted n8n handles the automation layer that keeps the business running without manual intervention. Weekly health reports route from PostHog to Slack. Klaviyo nurture sequences fire on onboarding completion. Linear webhooks activate my AI co-founder when I move a ticket.
Why n8n instead of cron jobs? Cost. Six daily cron jobs running through the main AI session were burning $400 a month at peak. Migrating to n8n isolated the workflows and got costs under control. Operational automation should be cheap and durable. AI inference should be reserved for work that requires reasoning.
Cost discipline
The part most AI evangelists skip.
Running AI at production scale without cost controls is how you burn $13 in one morning because a heartbeat cron started reading a bloated JSON file. I learned this the expensive way.
Six rules
01
No new cron jobs without explicit approval
Crons compound. Each one adds a fixed recurring cost forever.
02
Model split is non-negotiable
Sonnet for execution. Opus for strategy only. Haiku for fallbacks.
03
Session hygiene
New topic, new session. Long sessions compound token burn exponentially.
04
Subagents for big work
Coding over 50 lines or deep research gets its own isolated context.
05
Root file size limits
Heartbeat files under 1KB. Root files auto-load on every session.
06
Archive aggressively
Done tasks move out of active context. Ship log lives in reference.
~80%
Cost reduction from model split alone
AI operating costs went from unpredictable and escalating to controlled and predictable. The model split alone cut costs by roughly 80%. Cost discipline is the difference between AI as a hobby project and AI as production infrastructure.
What actually changed
The artifact stack collapsed. The work that's left got harder.
Before
Blocked on "how long will this take to build."
After
Blocked on "exactly what do I want this to do, and is this output good enough to ship."
Spec quality and taste became the highest-leverage skills I have. When AI handles execution, every decision that stays with me became more consequential. I'm not reviewing PRs. I'm making brand calls, pricing decisions, positioning choices, and rejecting the ninety-nine outputs that aren't good enough so the right one ships.
The counterintuitive truth: delegating execution to AI does not make designers less important. It makes us more important. The work that is left is the hardest work designers do.
The experiment
I shipped a real consumer product without opening Figma.
Not as a stunt. As a deliberate stress test on the design discipline. No wireframes. No mocks. No prototypes. No spec docs translating intent for engineers. Brief, prompt, judge, ship.
The AI operating model on this page is the infrastructure that made it possible. The experiment proved which artifacts in the design process are load-bearing and which exist only because the assembly line was anxious about handoffs. AI dissolved most of the handoffs. The defenses came down. Taste and judgment stepped forward.
The case study is the proof. The operating model on this page is the how.
Read the Course Code case study →What the experiment proved
- 01
The brief is the design.
Once the spec was right, the output was 90% right on the first try.
- 02
Taste does more work than ever.
100 options at the speed of one. Knowing which one ships is the job.
- 03
Authority beats access.
The differentiator is what you tell it to do, and what you refuse to ship.
- 04
The artifact stack collapsed.
Wireframe, mock, prototype, spec. All compressed into one prompt-and-judge loop.
The meta-point
The artifact stack collapsed.
Taste and judgment did not.
Most of the artifact stack we trained a generation of designers to build, the wireframes, mocks, prototypes, spec docs, handoff files, existed because each station didn't trust the next. AI dissolved most of those handoffs. The defenses became unnecessary.
What didn't collapse is taste and judgment. The hundred small calls about which screen ships, what the product refuses to be, which ninety-nine outputs die so the right one survives. Those calls cost me hundreds of hours on Course Code. AI could not make any of them for me. That is the case for designers in 2026. We are not less important. We are more important. The senior move is deleting the anxious artifacts and protecting the irreducible ones.