classify-info

Classify Info

Classify information into the right persistence mechanism: Hook, MagicDocs, Skill, CLAUDE.md, Memory, or None.

Decision tree

Actionable? Can Claude behave differently — in this project or any of the user's projects? No → None.

Fully mechanical? Specific command, no judgment, same every time → Hook. If the intent is mechanical but the exact command isn't specified, it's still a Hook — help fill in the command.

Describes or prescribes? Descriptions of architecture, data flow, how components connect → MagicDocs. "The frontend uses React context, not Redux" describes what the system is, even though it mentions what not to use. Prescriptions continue below.

Has steps? Multi-step procedures → Skill. Single constraints continue below.

Always relevant? Would Claude make a wrong assumption without this on its very first action in any session? Yes → CLAUDE.md. No → Memory.

Two common errors at this boundary:

• "Always true" ≠ "always relevant." "User prefers bundled PRs" is always true but not relevant when fixing a CSS bug → Memory. "Don't mock DB in integration tests" is always true but only matters when writing tests → Memory. "User finds excessive comments patronizing" is always true but many sessions are research/debugging/discussion without writing code → Memory.
• Ubiquitous actions are always relevant. "Make sure tests pass before committing" sounds topic-specific but committing happens most sessions → CLAUDE.md. "Deploy to Fly.io, not AWS" shapes config, CI, env vars broadly → CLAUDE.md.

Routing

Hook and MagicDocs are always project-scoped. For Skill, CLAUDE.md, and Memory, ask: does this follow the code or follow the person?

Classification	Action
Hook	Provide hook config for .claude/settings.json
MagicDocs	docs/magic/ if available; CLAUDE.md with advisory if not
Skill	REQUIRED SUB-SKILL: Use superpowers:writing-skills
CLAUDE.md	Project CLAUDE.md, subdirectory, or ~/.claude/CLAUDE.md
Memory	Project-scoped or ~/.claude/memory/
None	Explain why it's not actionable