AI-augmented development, Ruby on Rails, and engineering leadership
Claude Fable 5 ran a nonprofit's club-management system through Spec Kit's full round trip, twice, unattended, inside a free access window that would otherwise cost about $85. A cheaper model, picked for being the best fit for the language, builds it this weekend.
Token cost in a Claude Code swarm is a budgeting problem, not an intelligence problem. Three levers, model routing, progressive disclosure, and context packets, keep long migrations affordable and fast.
A multi-agent RSpec to Minitest swarm kept relearning the same conventions. Moving validation from prompts to regex-enforced gates turned every failure into a permanent fix.
You make a decision in February. By April, three people re-debate it because the context lived in a Slack thread. I built a Knowledge OS on Claude Code where capturing knowledge is a side effect of work. After 2.5 months: 188 artifacts, 2,500+ docs in semantic search, decision chains that compound.
764 Claude sessions across ~259 unique files over 16 working days. The orchestrator was a bash while loop. The real engineering was in layering error handling: a generation loop, a fix loop, a cleanup orchestrator, and a human catching what all three missed.
RSpec model specs: 6 minutes for 3,780 examples. Minitest model tests: 66 seconds for 9,835 tests. Both parallel. I built a 6-gate multi-agent pipeline in Claude Code to make the migration work.
I built a multi-agent system that applies mathematical testing techniques to 60+ RSpec models at scale. The architecture worked. The tests were rigorous. Then the results revealed the bottleneck was never test quality.
We used TestProf to diagnose a Rails monolith where factories consumed 95% of test time. Factory refactoring and let_it_be delivered 50-95% speedups on the worst specs. The full suite improved by 14%. The gap between those two numbers explains why manual optimization hits a ceiling.