The future of work isn't about prompting better — it's about teaching AI to work while you're offline.
The Summary
- Allie K. Miller, ex-AWS ML head and Time's top 100 AI influencer, runs her entire operation through multiple instances of Claude Code, Anthropic's agentic coding system with filesystem access
- She uses Claude's Skills feature to teach repeatable workflows: overnight email triage, calendar optimization, and auto-generated social content from video edits
- Her operating principle: "By the time you wake up, your AI should have already been working for you for hours"
The Signal
Miller isn't chatting with AI. She's delegating to it. The distinction matters. Through her consultancy Open Machine, she advises OpenAI, Google, Anthropic, and Warner Bros on AI adoption. With 1.6 million LinkedIn followers, she's not just talking about agents — she's living inside an agent-first workflow that most knowledge workers won't see for another two years.
The technical setup: multiple Claude Code terminals running simultaneously, each with filesystem access. That last part is key. These aren't sandboxed chatbots. They're reaching into her actual files, reading folders, triggering automations based on file events. When she drops a CapCut export into a specific folder, Claude generates the transcript, writes the social post, and creates the thumbnail. No human in the loop.
"By the time you wake up, your AI should have already been working for you for hours."
The Skills feature is where this gets interesting. Miller is training Claude on multistep processes, then letting it repeat them autonomously. Morning briefing that summarizes urgent emails and scans her calendar for recharge windows. When it spots back-to-back client calls, it blocks deep work time the next day. This isn't reactive. It's anticipatory.
Key workflow examples Miller has automated:
- Overnight email triage with urgency scoring and summaries
- Calendar analysis with proactive scheduling recommendations
- Video-to-social pipeline (export → transcript → post → thumbnail)
What Miller is describing is the shift from prompting to programming-by-demonstration. You don't write code. You show the agent the workflow once, maybe twice. It captures the pattern. Then it runs that pattern every time the trigger condition is met. This is closer to how you'd train a junior employee than how you'd use a tool.
The implication for knowledge work is stark. Miller's operation isn't augmented by AI. It's orchestrated by it. The human becomes the editor, the strategist, the person who reviews what got built overnight. If someone with Miller's profile and workload can collapse entire categories of admin work into background processes, the question isn't whether this scales. It's how fast.
The Implication
Start thinking in workflows, not tasks. The people who win the agent economy won't be the best prompters. They'll be the ones who can articulate their work as repeatable processes and teach agents to execute them autonomously. Miller's setup is advanced, but the principle scales down. Pick one repetitive multistep process you do weekly. Document it. Find the tool that can learn it. That's your entry point.
Watch what happens when agents get filesystem access and event-based triggers. That's when they stop being assistants and start being infrastructure. The gap between what Miller is doing today and what's available to the average knowledge worker is closing fast. The question is whether you'll be ready when your work can run while you sleep.