Build AI Second Brain: 36 Proactive Claude Agents

Agentic AI Delivers 2-10x Productivity Over Chatbots

Alli Miller, AI advisor to OpenAI, Google, and Anthropic, contrasts pre-agentic AI (20-30% gains from Q&A synthesis) with today's action-taking systems. Two years ago, AI required manual follow-up; now, her 36 proactive workflows with ~100 agents (28 master agents spawning ~50 sub-agents) handle hours of delegated work autonomously. Productivity jumps 2-10x per task, as agents schedule runs via Claude's tools, operating while she sleeps, walks her dog, or socializes.

"Depending on the task is anywhere between like 2x and 10x," Miller says, emphasizing agents as true delegates versus passive assistants. She routes outputs to email folders: Friday recaps scrape Gmail for unread urgent emails over 5 days, rank by urgency, draft replies, suggest team delegations, and add reminders. Morning briefings compile industry news, local events, weather, and meeting prep triggers hours before she wakes.

Host Marina notes the gap: someone finishing the tutorial and ingesting files into Claude gains a massive edge over non-adopters within a year.

Complain to Claude: Zero-Code Workflow Discovery

Miller's entry point for any automation: voice or text complaints to Claude. Stressed before client calls? Need umbrella alerts or deep work blocks? "The best first step to figure out what Claude should code to help you is just to complain," she advises. Humans excel at venting; Claude iterates solutions in real-time, proposing proactive agents without coding.

Live demo: Marina rambles a prompt for a 6 AM San Francisco morning brief—no calendar/email access yet—for Apple TV exec news (top 3 by impressiveness/buzz), 'game changer'-hyped AI stories, weather outfit advice, and 3 fun events in 4 days. Claude asks clarifying questions (time, format), then builds via its Skill Creator: reads instructions, plans 6 steps (research, summarize, schedule), and delivers a Word doc sample.

"All humans know how to complain. It's the joy that you get from having your complaint faced with not just like emotional validation... but like at a certain point, I don't want to be validated. I want that problem actually to be solved."

For sophistication, enable Claude's 'ask user questions' skill: it interviews for details (e.g., studio setup with mics, water, furniture), plans, then executes. No prompt engineering needed—rambling captures nuance better than concise inputs.

Claude's Four Versions and Action Layers

Miller breaks down Claude's ecosystem for escalating agency:

Web App: Basic chats, internet browsing, Notion/Gmail connectors. Great for retrieval, weak on actions.
Claude Co-work: Points at desktop files, creates Google Docs, runs code for APIs (Gmail, Fireflies, Granola). Business-focused agentic platform.
Claude Code: Full control for custom software, scheduling, local actions.
Chrome Extension: Automates browser tasks, e.g., collaging kid photos on Walgreens site by controlling mouse/keyboard.

All support skills in first three. Background code (e.g., API pulls) runs invisibly; users describe needs in natural language. Scheduling in Claude Code/Co-work ensures proactivity—no daily manual kicks.

Marina demos voice prompting in Claude Chat/Co-work; Miller confirms skills migrate across providers (Perplexity, ChatGPT) via folder uploads.

Skills as Modular Toolbox for Reusability

Skills are long prompts + logic in folders with examples/resources (e.g., CSVs of past social performance). Claude's toolbox analogy: pick existing (hammer for nails) or build new (wire cutters). Skill Creator automates this—ask Claude to interview, plan, code.

Examples:

Tone/brand voice for newsletters/LinkedIn.
Anti-AI language remover.
Survey data to action items.
Social media: post scripts, guest selection, performance analysis.

Embed skills in workflows: morning brief uses LinkedIn voice + DocX writer + scheduler. Modular for agent-to-agent sharing: LinkedIn agent passes anti-AI skill to Twitter agent.

"Agents teaching other agents new skills and being able to have these modular skills that I can throw over... there is going to be a lot of agent to agent sharing."

Doubt a task needs a skill? Ask Claude: describe your day, get 3 suggestions. Push back if needed—"emotional fortitude" required.

Four AI Models: Delegate to Teammate

Miller's framework classifies agents:

Microtasker: Simple tasks.
Companion: Q&A buddy.
Delegate: Assigned work (morning brief).
Teammate: Proactive, team-scale (e.g., Jira analysis for project progress, shared briefings).

Top users treat AI as "first class teammate," not "intern." "I actually get pretty annoyed when I hear people say, 'Oh, AI is an intern.' I'm like, 'What intern has PhD level intelligence, the ability to read the entire internet?'"

Enterprises hoard super-user gains; teammates reduce friction for laggards, uplifting departments. SMBs/solos: use for onboarding skeptics.

Host's Miro AI plug highlights context challenges—AI needs team knowledge (strategies, tasks). Canvas-as-prompt grounds agents in files (Alli's LinkedIn, newsletters), spawning sidekicks for research, flows for themes (AI setups, predictions, advice).

Mindset Shift: Automate Repetition, Scale Teams

Impact in a month: faster tasks, mindset for business applications (marketing, sales, products). Less terror amid AI pace—see direction via proactivity.

No tech skills needed; APIs invisible. Start small: daily news → full systems. Gap to non-adopters: massive, as agents compound.

Miller's photo sync gripe → Claude's Drive folder + classification + team email solution shows iteration joy.

Key Takeaways

Complain to Claude about pains (umbrella alerts, deep work); it proposes/codes proactive agents—no coding required.
Build morning briefings: industry news, weather, events via rambling prompts; schedule for 6 AM delivery.
Use 'ask user questions' skill for interviews leading to custom setups like studio gear or email recaps.
Create skills as folders (prompts + examples); modular across Claude versions/providers for tone, brand, analysis.
Schedule workflows in Claude Code/Co-work for 24/7 runs (e.g., Friday Gmail urgency ranks/drafts).
Classify agents: delegate (personal) to teammate (team Jira, shared briefs) for 2-10x gains.
Migrate skills to Perplexity/ChatGPT; push back on refusals with examples.
Automate repetition: daily competitor checks, meeting prep triggers.
Treat AI as PhD teammate, not intern—share for team uplift.
Demo files into tools like Miro Canvas for grounded research agents.