This AI Second Brain Debugs Code While You Sleep

Hamish Davidson wakes up to code he didn't write. Not in the "oh no, I got hacked" way—in the "my AI assistant shipped three bug fixes and redesigned my pricing page" way.

Davidson, who runs the SaaS product Harbor, built what he calls a "second brain" using Claude Code, Obsidian, and Anthropic's Opus 4.6 model. The system runs autonomously while he sleeps, scanning his application for issues through browser audits and PostHog analytics, pulling tasks from Linear (his project management tool), fixing the code, and pushing changes to his development environment. No human involvement required until it's time to review and deploy.

The demo footage is honestly kind of surreal. Davidson shows his Vercel deployment log filled with commits like "Three surgical fixes driven by a fresh browser audit, plus 7-day PostHog scan." Another: "Remove duplicate FAQs from home page." The system identified a pricing display issue (showing "$24.5/month" instead of the cleaner "$24/month") and fixed it on its own. It even autonomously redesigned his pricing page after determining—through its own analysis—that conversion rates weren't where they should be.

What "Skills" Actually Mean

The architecture here matters. Davidson's system is built on Cole Medin's Second Brain Starter Pack, an open-source framework that uses the Claude Agent SDK to create what it calls "skills"—basically automated workflows that run on schedules you define.

One skill he demonstrates is the "product loop," which conducts a full product management audit of Harbor. When he triggers it (though it can also run automatically), the system announces: "Running the product loop, let me gather all signals to pick up the highest impact item." Then it actually goes and finds things to fix.

Another skill, which Davidson calls "churn autopsy," connects to Stripe to identify recently churned customers, then attempts to synthesize why they left by cross-referencing usage data from PostHog. "If you know anything about management, this is an absolute game-changer," he says. And yeah—automated churn analysis that doesn't require a data analyst to manually pull reports? That changes the economics of customer research pretty significantly.

The system maintains what it calls "known non-issues"—essentially a list of false positives it's learned to ignore. This is the difference between a tool that floods you with noise and one that actually becomes useful over time.

The Setup (And The Gaps)

Davidson walks through his installation process, which he admits he modified from Medin's original approach. Instead of manually filling out a requirements file, he has Claude Code interview him, asking questions one at a time: What's your role? What platforms do you use daily? What should your second brain never do?

That last question matters. Davidson's explicit instructions: never push to main, never push to production, never delete anything. The system pushes to development environments only, which means there's still a human review step before anything touches real users.

The integrations are where things get technical. You need API keys for Linear, OAuth credentials for Gmail, PostHog configuration, Stripe access—essentially, you're giving an AI system keys to your entire business stack. Davidson uses Python files for most integrations, though he mentions he's "starting to use MCPs more and more." (MCP, or Model Context Protocol, is Anthropic's standardized way of connecting AI systems to external tools.)

Setup takes 30 minutes to an hour, he estimates, after which the system builds itself. Then you wait. "You do also have to give it time, like it needs like a week or two, to be honest with you, to actually start understanding what you want every single day," Davidson explains.

What It Actually Ships

The most concrete example: Davidson says he "implemented the entire agency dashboard in under 24 hours using this system." Not that the AI wrote all the code—but that the workflow of identifying requirements, writing code, testing, and pushing to dev happened autonomously enough that a feature that would typically take days or weeks got to the review stage in one day.

Other fixes the system has shipped include:

• Removing duplicate FAQ sections it found during a browser audit
• Fixing a Grammarly integration bug identified through PostHog error tracking
• Resolving a gradient background issue that a team member had logged in Linear

Each of these represents the system successfully completing the full loop: detect problem → verify it's actually a problem → write fix → test → push to dev.

The Honest Limitations

Davidson doesn't oversell this. Multiple times he notes: "There are some holes in this, I have to say, and I think it's still not quite there yet."

The system needs constant refinement. After the initial setup, he recommends telling it: "Okay, that's great, but now let's really make this way better and more proactive." It's not plug-and-play magic—it's a tool that requires ongoing calibration to your specific needs.

There's also the trust question. Even with guardrails preventing production deployments, you're giving an AI system the ability to modify your codebase. That's a significant leap from "AI writes code suggestions you review" to "AI commits code to your repository while you sleep." The fact that everything goes through dev first is crucial, but you're still betting your development velocity on the system not creating more problems than it solves.

And then there's the philosophical tension: "AI doesn't know what you want. Only you know what you want," Davidson says. "So you have to sit here and make sure that this is your second brain in the sense that it has everything it needs to be your second brain. It can't be your second brain if it only has half of the information."

What This Actually Represents

This isn't AGI. It's not even close. What it is: a specific implementation of AI agents that moves the automation boundary in software development. Previously, AI could suggest code. Then it could write functions. Now it can complete the full cycle of bug detection, triage, fixing, and deployment to a staging environment.

The comparison Davidson makes is to OpenClaw (another autonomous coding system), but he argues this is "safer, and better" because it's built on official APIs—the Claude Code CLI and Claude Agent SDK—rather than OAuth hacks.

What makes this particularly interesting is the integration layer. It's not just an AI that writes code in isolation. It's an AI that understands your business context through Stripe, your user behavior through PostHog, your roadmap through Linear, and your infrastructure through Vercel. That contextual awareness is what enables it to make decisions about what to fix, not just how to fix it.

Davidson gives credit throughout to Cole Medin, whose Second Brain Starter Pack is open source on GitHub. The fact that this is buildable by other developers—not a proprietary black box—means we're probably about to see a lot of experiments in this direction.

Whether those experiments make developers more productive or just create new categories of bugs to debug is the question we're all about to find out together.

—Zara Chen