Stop over-constraining your AI agents: prune rules, keep conventions.

Boris Churney, creator of Claude Code, sparked confusion by suggesting teams delete their entire AI layer every six months. The real message is different: run ablation tests by removing all instructions, then adding them back line by line to measure actual impact. This helps eliminate rules that over-constrain modern LLMs.

The critical insight is that better models handle reasoning on their own, so many rules just fix gaps in older models. However, project-specific conventions—style guides, test patterns—still matter. The creator tested this on a 1000-line ruleset and found stripped-down agents matched full performance on architectural decisions but failed on custom conventions.

The practical challenge is cost. Ablation runs are token-heavy and time-consuming. Churney's team runs two-week sessions, which is impractical for most teams. The creator recommends focusing ablation efforts on global rules rather than skills or sub-agents.

In summary, prune aggressively for reasoning-related rules but keep conventions-specific ones. Do this regularly but not every six months. The real bottleneck isn't model capability—it's teaching models how your team works.