Better Models Make Worse Tools: Claude’s RL Harness Teaches It to Invent Tool Schema Keys

“The better the model gets, the worse it becomes at using your tools—Claude is now inventing JSON keys because its training forgave too much slop.”

8.3Weirdness

Why It Matters

As models get "better" via targeted RL in one harness (Claude Code), they degrade on other tool schemas - schemas are no longer neutral. Concrete transcripts show harness quirks dominating builder experience.

Evidence

Armin Ronacher’s lucumr.pocoo.org blog post (July 4, 2026, HN top), concrete malformed JSON examples from Pi editor calls (invented keys like "requireUnique", "in_file"); tied to Claude Code/Opus/Sonnet training quirks.

Signal Read

Novelty: 8Receipts: 9Story voltage: 8Heat: 8

Source Trail

Daily scan: 2026-07-05

  • No public source URL captured yet.