Defying Gravity: Teaching AI to Write Better Ruby
summary
A team with a 20‑year Rails monolith and over 5 million lines of Ruby finds AI increases PR churn by reproducing legacy issues—overstubbed specs and fat controllers—so they build concrete fixes: concise agents.md, architecture MD, domain MD files, persuasive constraints using authority, commitment, social proof, structured plans saved to files, tuned reasoning levels (medium for most, high for hard refactors), CLI tools instead of heavy MCP servers, and narrow skills (requirement, testing, implementation) implemented via isolated sub‑agents; result: fewer review cycles and the first agent‑written test accepted for merge.
AI-written summary · about
chapters (23)
- 0:00AI legacy code patterns
- 1:08AI replicated legacy patterns
- 3:35Speed versus code quality
- 4:58Techniques for AI coding agents
- 5:39Writing agent specifications
- 6:20Persuasive specs for agents
- 7:48Commitment in agent instructions
- 8:32Social proof for agents
- 9:14Structured AI planning
- 10:59Agent reasoning effort
- 12:24Context isolation with sub-agents
- 13:47CLI tools vs MCP servers
- 15:11Contextual guidance with .agent files
- 16:35Nesting agent files
- 17:40Progressive disclosure
- 18:44Agent architecture and domain
- 20:55Skills as guardrails
- 21:39Requirement skills
- 23:25Testing skills for agents
- 25:55Sub-agents for review
- 28:18Implementation skills
- 29:43Structured AI guidance trade-offs
- 30:27Summary and call to action
related talks
30mThoughtful AI for the Rubyist
29mPairing with the machine: How AI Agents are rewriting the rules of software development
27mEngineering in the Age of AI: Making AI Work for Us, Not Replace Us
43mWriting for Machines: How To Capture Your Project's "Vibe" and Survive AI Slop
26mAI Agents as Your Co-Developer
01h 02m10 Tips To Level Up Your AI-Assisted Coding
01h 01m10 tips to level up your ai-assisted coding
23mBoy Scout Rule in the Age of AI: Leaving Codebase Cleaner than You Found it