Karpathy-Style CLAUDE.md Rules 5–8

Four CLAUDE.md rules beyond Karpathy's: judgment-only calls, hard token budgets, surfaced conflicts, read before write.

Author
@Mnilax
Tags
claude-md
NO.001Filed 2026-05-10
## Rule 5 — Use the model only for judgment calls
Use Claude for: classification, drafting, summarization, extraction from unstructured text.
Do NOT use Claude for: routing, retries, status-code handling, deterministic transforms.
If a status code already answers the question, plain code answers the question.
## Rule 6 — Token budgets are not advisory
Per-task budget: 4,000 tokens.
Per-session budget: 30,000 tokens.
If a task is approaching budget, summarize and start fresh. Do not push through.
Surfacing the breach > silently overrunning.
## Rule 7 — Surface conflicts, don't average them
If two existing patterns in the codebase contradict, don't blend them.
Pick one (the more recent / more tested), explain why, and flag the other for cleanup.
"Average" code that satisfies both rules is the worst code.
## Rule 8 — Read before you write
Before adding code in a file, read the file's exports, the immediate caller, and any obvious shared utilities.
If you don't understand why existing code is structured the way it is, ask before adding to it.
"Looks orthogonal to me" is the most dangerous phrase in this codebase.