Convention Over Hallucination: Harness Engineering for AI-Powered...
summary
Claims that making AI output trustworthy by default requires a harness with four pillars: instruct (shared architecture docs and a feature memory using the mealas plugin and a Milo-style repo with ~50,000 stars), constraint (RuboCop-style cops turned into actionable lint messages), workflow (a Compound Engineering plugin encoding skills and solution folders), and verification (agent-driven end-to-end browser checks). Dropped into a 7-year-old codebase, the RuboCop harness requires grandfathering, CI hard-fails on convention breaks, and constraints prove the highest-leverage investment after months of iteration.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 3:57AI code generation journey
- 6:05Convention over configuration
- 9:38Enforcing conventions
- 13:34Workflow automation
- 15:41Verification agents
- 17:28Shift to system design
- 19:39Actionable steps
related talks
29mBuilding Trust in AI Agents
31mEngineering the harness
25mWhen Your "New Teammate" Hallucinates: Observability as the Safety Net for AI Code
50mPanel: Beyond The Vibe: Code Quality First
48mNever Trust a Monkey: The Chasm, the Craft & the Chain of AI-Assisted Code
46mHarness Engineering: How to Build Software When Humans Steer, Agents Execute
30mWhen AI Takes Over: DevOps Nightmare Unleashed
50mFrom Prompt Hacking to Architectural Determinism: Engineering Reliable GenAI Systems