Caching coverage is 30% — room to improve.

Keep system prompts stable; avoid one-off prompt edits that bust cache.

System prompt is very large (~2552 words) — heavy fixed input every request.

Move rarely-used rules to docs and link from a short root CLAUDE.md / AGENTS.md.

12 tool calls/session on average — large tool JSON payloads inflate input tokens.

Narrow tool surface, summarize tool results before re-prompting, and use cheaper models for tool routing.

Model choice matches typical coding workloads.

Use a tiered model policy: router for simple edits, premium model for refactors.

System prompt is 2552 words — dominates every request.

Split into scoped skills files; keep root prompt under a few hundred words.

150K-line repo with full-file reads — agents ingest thousands of irrelevant lines per task.

Use jCodeMunch or a tree-sitter retrieval tool to read only the symbols agents need — the vendor reports up to 95% token reduction.