Real claude -p cost: $0.27〜0.47 before any documents

The company's first lesson came from content, not volume. When its CLAUDE.md passed 1,000 lines, the rules contradicted each other and the AI began acting against instructions, so the main file was kept under 200 lines and details split into .claude/agents/. Its conclusion is that fitting inside a 1M-token context and being used correctly are separate questions: if old and new decisions both sit in the context, the AI can follow either.

That is why its first job for whole-document reading is checking for contradictions, not answering questions. Concrete tasks include sweeping the decision log, rules, and department STATE files for conflicts, auditing whether descriptions of shut-down businesses still linger in other documents, and listing gaps between the quarterly roadmap and actual progress—work where you cannot know in advance which file to open. What it never delegates is the decision of which document prevails; the CEO decides, and one line goes into the decision log.

The wrapper it substitutes for bare claude -p adds --output-format json so each response carries total_cost_usd, pipes it into record-cost.py to accumulate in cost-tracker.json, and sets --max-budget-usd 3 as a per-run cap. It warns against --bare, which does not read authentication and fails unattended runs. Prompts pass file paths rather than document contents, and the most effective habit was not the new feature but measuring one call before deciding to adopt.