When 100% compliance is required (financial, safety, identity), the answer is always programmatic enforcement — never stronger prompts or more examples.

If a workflow blocks a human or requires a response in under an hour, it cannot use Batch API. Batch has no latency SLA — up to 24h.

Subagents inherit nothing from the coordinator. Every piece of needed context must be explicitly included in their prompt. No shared sessions, no automatic inheritance.

Only 3 valid triggers: explicit human request, policy gap, inability to progress. Sentiment scores, confidence scores, and complexity alone are NOT valid escalation triggers.

~/.claude/CLAUDE.md is user-level and NOT shared via git. Team conventions belong in .claude/CLAUDE.md (project-level, version-controlled).

JSON schema + tool_use eliminates syntax errors (wrong type, missing field). It cannot prevent semantic errors — correct type placed in wrong field.

In long context windows, models reliably process start and end but miss middle sections. Put critical info at start/end. Use section headers for structure.

Use Plan mode for: large-scale changes, multiple valid approaches, architectural decisions, 10+ files. Direct execution for clear, scoped, well-defined tasks.