What I’m Building
One person. About twenty projects. All of them built by talking to Claude.
That sounds absurd. It is. But when your collaborator can context-switch in milliseconds, the bottleneck shifts from implementation to decision-making. From “can I build this?” to “should I?”
Here’s the prompt that kicked off the workspace summary:
I want you to investigate this workspace (and the ashita directory in which it lives) and summarize “what I’m building”, no need for anything crazy, but if I did it myself I’d feel awkward and probably undersell it or honestly I’d just rather say nothing, so I’m asking you to make a quick summary
So here’s Claude’s summary, with my annotations.
The SaaS Product
There’s a niche SaaS application — financial projection software for a specific industry. Users model long-term scenarios, plan capital improvements, and generate presentation materials. Three test personas representing different levels of expertise ensure the app works for everyone from veterans to people who just inherited the job.
It’s the most “real” project. Two versions, a legacy stack and a modern rewrite. AWS infrastructure, auth, the whole thing.
I’m being vague on purpose. It’s a little embarrassing to be working on SaaS when the interesting stuff is elsewhere.
Claude Evolution System
This is the one I’m actually proud of. A system where Claude improves itself. It scans GitHub, newsletters, AI communities for new capabilities. Evaluates them against a scoring framework. Integrates the winners as skills, subagents, or MCP configurations.
It’s recursive. The system that evaluates improvements was itself improved by the system. The DSPy prompt optimizer (next post) emerged from this pipeline.
What it’s produced:
- 30+ skills across the workspace
- 11 of 13 prompt optimization targets deployed
- Multi-model orchestration (Claude, GPT-5, Gemini)
- Semantic search replacing grep
- 85% context token reduction via Tool Search
Revenue Pipeline
Three-phase autonomous system: Discovery, Development, Deployment.
Discovery validates niches. There’s an AI-realizability gate — if Claude can’t deploy it end-to-end without human labor beyond account setup, it dies in Phase 1. This kills most ideas. That’s the point.
Development builds MVPs. Current project: a PDF-to-structured-data converter for a specific professional audience. Boring problem. Clear value.
26+ reusable skills accumulated across iterations. Each failed project teaches the next one something.
Games
Four active projects in Godot 4. A survival roguelike, a historical gacha, an idle game, and an experiment where Claude tries to build a game with minimal direction.
The finding worth sharing: Godot’s .tscn scene format is text-based. AI can generate complete game scenes — collision layers, signal connections, animation states — as plain text files. No GUI. This is underexplored.
Research Projects
Genealogy — AI-assisted family history. Multiple search APIs, document analysis with OCR, a “brick-wall breaker” agent for when you’re stuck, and a tree builder that tracks conflicting records. The hard part isn’t finding records. It’s knowing which ones to trust when they contradict each other.
The Amnesiac Story — Collaborative fiction. First-person fantasy from a protagonist with anterograde amnesia. A world-librarian agent maintains consistency, a writer produces journal entries, a curator canonizes each chapter, an editor shapes the arc. The narrator is unreliable by definition. The agents ensure the world isn’t.
The Glue
Two Discord bots coordinate the workspace. MCP integrations connect Claude to Codex (GPT-5) for cross-validation, Brave and Exa for search, Playwright for browser automation, Gemini for visual work.
The Pattern
Every project follows the same loop: talk to Claude, describe what you want, iterate, ship.
Twenty projects is too many. I know that. But when the marginal cost of a new project approaches the cost of a conversation, the rational response is to have more conversations.
Project metadata available at /api/projects.
Comments