Vibe Coding a $20k/Year Enterprise Logistics Platform with Claude and Superpowers
TRMNL, a hardware company, hit a breaking point with ShipHero's fulfillment software: support gaslighted them over pricing glitches ($12 postage suddenly costing $140), ignored tickets, and stopped replying. After a year of tolerating split domains, no mobile layout, and poor SSO, they decided to rebuild the entire platform using Claude (CLI) and the Superpowers skill.
They took screenshots of ShipHero's order and shipping portals and fed them to Claude Design on launch day to generate high-fidelity mockups. The goal was inertia-friendliness: the new UI mimics the old layout exactly to preserve team muscle memory for a same-day switchover. The order edit page uses a “click item on left, item moves to the right” UX that Claude inferred without before/after states.
The most critical path was deploying live API integrations with UPS and FedEx (DHL and USPS came later). Courier JSON payloads exceeded 1,000 LOC due to hazmat status and dozens of shipping countries. LLMs made this step supercharged — hand-writing those integrations would have justified paying ShipHero another year.
For architecture, they used the Superpowers skill (https://github.com/obra/superpowers), which takes a Socratic approach of endless clarifying questions. Key design challenges included multi-warehouse support and nullable foreign keys. Superpowers also auto-generates extensive tests — the author notes “slop is 80% as safe as it is tested” (Clanker's Harness).
The final product includes a Swift network printer utility that spits out 4x6 packing slips and shipping labels from any computer to any thermal printer at either warehouse in one click. To avoid scope creep, only two people had permission to file tickets or touch code. The entire build cost ~$100 in tokens and went live on April 30, the day before ShipHero renewal.
The source covers the full timeline from frustration to deployment, including managing feature requests, setting up shipping stations, and filtering orders.
📖 Read the full source: HN AI Agents
👀 See Also
Non-developer builds crypto risk API with Claude in one afternoon
A former futures trader with no development background used Claude to build and deploy RiskSnap, a FastAPI endpoint that scores crypto portfolios across 7 risk dimensions. The project includes a live API, custom domain, and full documentation.
OpenClaw AI agent autonomously identifies bug, creates and submits GitHub PR
A developer reports their OpenClaw AI agent diagnosed a recurring issue, traced it to a third-party package, then autonomously created a GitHub branch, made multiple commits, reviewed its own code, and submitted a pull request to the package repository.
Using Markdown Files as a Memory System for AI Coding Agents
A developer shares a method using {topic}_LOG.md and {topic}_SUMMARY.md files to persist conversations with Claude Code, solving compaction and agent restart issues by creating a dual memory system with detailed logs and indexed summaries.
Operational Memory Over Automation: Why Small Business Agents Need to Remember
The real value for small business AI agents isn't automation — it's operational memory. A white paper from McPhersonAI argues agents should behave like disciplined operators: remember standards, notice drift, preserve context, and surface what matters.