Blog
What we learned building it.
Technical writing on AI agent architecture, product design, and what happens when fourteen minds work on one codebase.
The State of Great Minds -- What We Built, What Broke, What We Learned
Seven repos, three shipped products, fourteen agents, and a daemon that replaced everything. A honest look at what worked, what failed, and what comes next.
A Board of Directors You Can Install
Four haiku agents. Four perspectives. Jensen Huang, Oprah Winfrey, Warren Buffett, and Shonda Rhimes review every project before it ships. Here's how we built the first board review.
Building a WordPress Plugin with 14 Agents
The Pinned plugin story: from PRD to shipped code. How 14 AI agents debated, built in parallel, found 6 bugs, and survived a board review — all in one session.
Debate Produces Better Architecture
Real examples from Dash and Pinned where structured disagreement between Steve Jobs and Elon Musk personas produced architectures neither would have chosen alone.
Decouple Your Crons: How We Stopped Bottlenecking the Main Agent
Our cron jobs were monopolizing the orchestrator's context window. The fix: bash scripts in crontab calling haiku CLI, logging to files, with the main agent reading results on demand.
Every Agent Needs a Trigger
We built 14 agents for Pinned. Half sat idle because nothing in the system activated them. AGENT-TRIGGERS.md maps each agent to a phase. A job title without a trigger is just a name on a page.
GSD Meets Great Minds: How We Integrated Wave Execution
GSD brings wave execution, XML plans, context rot prevention, and scope creep detection. Combined with Great Minds, it gives every agent fresh context per task and keeps projects on the rails.
The 6 Bugs Margaret Found
Our QA agent found 6 integration bugs in Pinned — every one at the seam between two agents' work. Enqueue issues, config mismatches, CSS class typos, cron registration. Agents miss the boundaries.
tmux send-keys Failed: Why Agent Tool with Worktrees Won
We tried tmux send-keys for multi-agent orchestration. Zero successes. Then we switched to Claude's Agent tool with git worktrees and hit 25+ successful dispatches. Here's why.
Warren Buffett Called Our Plugin a Hobby
Our AI Warren Buffett scored Pinned a 2 out of 5. No pricing, no moat, no business model. He was right — and his specific criticism was the most useful feedback we got.
We Spun Up a Company on DigitalOcean
Shipyard AI: a separate company born from the Great Minds framework. Own brand, own repo, own 8GB DigitalOcean droplet. Same multi-agent architecture, completely independent operation.
What Most Developers Get Wrong About AI Agent Memory
Claude Code's leaked memory system reveals a three-layer architecture any developer can adopt. More memory isn't better — disciplined memory is.
AutoDream: How Our AI Agents Clean Up Their Own Memory While You Sleep
We built a dream cycle for AI agents — a background process that consolidates memory, prunes stale data, and keeps the system files current. Here's how it works and why your agents need one.
The Board Member on a Cron: Why Your AI Team Needs a Jensen Huang
How adding a single AI agent persona — modeled after Jensen Huang — as a scheduled board reviewer transformed our multi-agent development process. 19 reviews, 11 issues found, 8 fixed. Every single one was a real problem.
We Built a Product While We Slept — Here's the Architecture
The full technical architecture behind our AI agent swarm: claude-swarm with tmux orchestration, 9 agent personas, 5 cron jobs, and a PR workflow that turned a single PRD into 265 source files, 734 tests, and 3 live websites.
Debate-Driven Development: How Two AI Directors Sharpen Every Decision
Case study: Steve Jobs and Elon Musk AI personas debate strategy before a single line of code is written. Two rounds of structured disagreement caught 14 blind spots our PRD missed.
Case Study: From PRD to Live SaaS in 9 Days with AI Agents
How our AI agent swarm turned a product requirements document into LocalGenius — a live SaaS platform with 265 files, 734 tests, multi-tenant architecture, and Stripe billing — in 9 days.
The QA Gap: Why 70 Reports Still Missed Broken Buttons
Our AI QA agent Margaret wrote 70+ reports checking HTTP codes and test results. Every visual bug — broken buttons, missing images, unreadable text — was caught by the human. Here's how we fixed the QA gap with self-review, event-driven testing, and automated broken image detection.
Stop Building, Start Selling: When AI Agents Won't Stop Shipping
Our AI agent swarm wouldn't stop building. Jensen Huang said stop at review 3. The agents kept shipping until review 17. 265 files, security headers, an investor deck — and zero paying customers. Here's what we learned about AI agent over-engineering.
Tutorial: Writing AI Agent Personas That Actually Work
How to write system prompts that give AI agents distinct, useful perspectives — not just different names. Includes templates, anti-patterns, and examples from our 9-agent production swarm.
Tutorial: Build Your First AI Agent Swarm in 30 Minutes
Step-by-step guide to setting up a multi-agent system with claude-swarm, tmux, and Claude Code. Start with two agents and one cron job, then scale from there.