Decision Tree: Which to Use?

Does correctness require guaranteed ordering, retries, a required artifact, or a hard stop?
├─ Yes → Use a HOOK, SCRIPT, CI JOB, or DYNAMIC WORKFLOW
│        Keep judgment in a skill only when the model must interpret a case.
│
└─ No → Is this reusable knowledge or an adaptable procedure?
        ├─ Yes → Use a SKILL
        │        Manual-only if timing or side effects require user control.
        │
        └─ No → Does this need isolated context or parallel work?
                ├─ Yes → Use an AGENT
                │        Example: code-reviewer, performance-auditor
                │
                └─ No → Just write it in CLAUDE.md as instructions

Starting heuristic, not a platform rule: content relevant to most sessions belongs in CLAUDE.md; procedures and references used occasionally belong in skills. Measure your own session mix before using a numeric threshold. Skill descriptions still consume context while visible, and the full body enters the conversation when invoked.

See also: Memory Loading Comparison for a broader decision tree covering all seven mechanisms (including Hooks, MCP, and CLAUDE.md vs rules). To automate detection of what belongs in each category, use cc-sessions discover, which applies this 20% threshold to your actual session history.