Skill Claude CodeCodex
Separates policy (what to decide) from mechanism (how to execute it). Use when a system has both a rules layer and a doing layer, or when business rules tangle with the code that runs them. Do not use to add layers a small tool does not need.
Skill Claude CodeCodex
Emits nothing when nothing surprising happened. Use when adding logging, progress messages, or success notifications, or when reviewing CLI output with banners and 'Done!' lines. Do not use to suppress errors or warnings — those are surprises.
Skill Claude CodeCodex
Chooses plain text over binary for configs, logs, IPC, and persisted state. Use when picking a serialization or reviewing a proposal for a custom binary format. Do not use where measured throughput or size requirements rule text out.
Skill Claude CodeCodex
Builds the thinnest real end-to-end path first, then fleshes it out. Use when starting a feature, integration, or system with unknown end-to-end behavior, or crossing a service boundary for the first time. Do not use for throwaway experiments.
Skill Claude CodeCodex
Writes or revises a SKILL.md so an agent triggers it at the right time and follows it reliably. Use when creating, editing, or reviewing a skill. Do not use for instructions that apply to every session — put those in AGENTS.md or CLAUDE.md.