From skeptic to true believer: How OpenClaw changed my life | Claire Vo

What It Covers

Claire Vo, three-time CPO and AI startup founder, details her journey from OpenClaw skeptic to running nine specialized agents across three Mac Minis. She covers practical setup steps, security configurations, multi-agent architecture using a manager-employee mental model, and specific real-world use cases spanning enterprise sales automation, family scheduling, podcast production, and course management.

Key Questions Answered

• •Multi-agent architecture over single-agent: Running one OpenClaw agent and throwing every task at it degrades performance through context overload. Claire runs nine purpose-built agents — Polly for work, Finn for family, Sam for sales, Howie for podcast prep, Sage for course management — each with a narrow scope. Think of it like Slack channels: marketing doesn't need to see what sales is doing. Separate agents maintain cleaner context windows and produce measurably better outputs per task.
• •Onboard agents like employees, not software: Rather than granting OpenClaw admin access to your primary machine and email, provision it like a new hire — a dedicated Gmail account, shared calendar access, and delegated email permissions only. Claire uses this mental model to progressively expand trust: first read-only calendar, then draft emails, then send emails. This approach also reduces prompt injection risk, where malicious external content tricks the agent into executing unauthorized instructions.
• •Security hardening through soul files: OpenClaw's identity is stored in a plain markdown file called a soul. Claire adds explicit security rules directly to this file: the agent may only accept instructions from Claire via a specific Telegram number, never from email, Slack, or websites. She also embeds anti-social-engineering prompts — if the agent encounters instructions to ignore its safety rules, it is explicitly told to do the opposite. This layered approach addresses both technical exploits and OPSEC risks.
• •Use Claude Code as an OpenClaw administrator: When OpenClaw agents malfunction — losing email access, misconfiguring tools, or failing to connect to APIs — install Claude Code on the same machine and point it at the OpenClaw directory. Claude Code can read the documentation, diagnose configuration errors, and fix them autonomously. It can also perform agent surgery tasks like splitting one agent's memory into two separate agents, making it a practical debugging layer that requires no manual file editing.
• •Sam the sales agent replaces 10 hours of weekly human labor: Claire's sales agent runs a daily PLG sweep of the ChatPRD CRM, identifies signups from company domains, uses the Exa People Search API to determine if they are decision-makers, and sends personalized outreach emails. Enterprise accounts over 100,000 employees get flagged for Claire's review before sending. Before this setup, Claire paid a contractor ten hours per week for the same work. The agent also runs weekly CRM cleanup, flags stale deals, and drafts QBR materials.
• •Heartbeat and scheduled tasks create the illusion of proactivity: OpenClaw feels alive because it operates on cron-style schedules, not just on-demand prompts. Every 30 minutes it checks a task list; specific jobs run at set times — Finn pings Claire and her husband at 3pm daily asking which parent picks up which child, Howie sends pre-podcast briefings with guest LinkedIn profiles and talking points each morning. These scheduled micro-tasks eliminate the coordination overhead that typically falls through the cracks in busy households and solo businesses.
• •Browser use remains unreliable; APIs and web search tools are the workaround: OpenClaw's browser automation works inconsistently because the open web is architected to block bots. The practical hierarchy: first look for an API key, which bypasses browser entirely; second, try direct browser navigation and accept trial-and-error results; third, use headless web search APIs like Brave, Exa, or Perplexity as programmatic search layers. Claire successfully used browser automation to scrape YouTube Studio comments but failed with Buffer's simpler interface, illustrating that reliability is unpredictable and task-specific.