| name | instruction-placement-framework |
| description | Decide where to place Claude Code instructions (CLAUDE.md, rules, skills, subagents, hooks, output styles, or appended system prompt) based on load timing, compaction behavior, context cost, and authority. |
Instructions
Use this skill when you are configuring Claude Code and want a consistent way to choose where an instruction should live.
-
Classify the instruction:
- Project facts that should always be in context → CLAUDE.md (root)
- Team or folder-specific conventions → CLAUDE.md (subdirectory) or path-scoped rules
- Specific constraints that must hold reliably → rules (prefer path-scoped)
- A repeatable procedure (runbook/checklist) → skill
- A side task that should run in isolation and return only a summary → subagent
- Deterministic automation or enforcement on lifecycle events → hook
- A major, persistent change in role or tone → output style
- One-off additive instructions for a single invocation → append-system-prompt
-
Check tradeoffs:
- If always-on instructions are growing, reduce context cost by using path scoping or skills.
- If you need deterministic enforcement, prefer hooks/permissions over prompts.
- If you don’t want intermediate results cluttering the main thread, use subagents.
Examples
- If you wrote a 30-line deployment procedure in CLAUDE.md, move it into a skill.
- If a rule only applies to
src/api/**, add paths: frontmatter so it loads only when relevant.
- If you need to block a dangerous tool call reliably, use a PreToolUse hook and exit code 2.
See: references/methods-table.md
See: examples/rules-paths-frontmatter.md
