Question 37
Domain 5A repo's CLAUDE.md has grown to 12,000 tokens over six months. Sessions feel slower to start, and the operator notices Claude sometimes parrots instructions back rather than acting on them. What is the most appropriate action?
Correct answer: B
Explanation
CLAUDE.md should stay focused on active operating instructions, because long context files slow startup and can cause the model to echo instructions instead of using them. Moving historical material into docs/reference/ and using @-includes for only the binding parts keeps the prompt lean while preserving access to background information.
Why each option is right or wrong
A. Switch the model to a smaller one to reduce per-token cost.
Changing to a smaller model lowers cost, but does not fix bloated instruction design.
B. Move historical sections into docs/reference/ and @-include only the parts that are still actively binding.
CLAUDE.md is meant to carry only currently binding operating instructions; once it has ballooned to 12,000 tokens, the prompt overhead itself becomes the problem, because every session must ingest that context before work begins. The appropriate remediation is to archive non-binding history in docs/reference/ and keep only the active sections in the root file via @-includes, so the model sees the enforceable instructions without paying the latency and attention-cost of the obsolete material.
C. Delete the file and re-derive context from each new session's user prompt.
Removing persistent guidance entirely loses stable operating rules and increases inconsistency across sessions.
D. Compress the file with an LLM-based summarizer and overwrite it in place.
Summarizing in place can discard nuance and mix active rules with historical context.