Save

Reviews the current session and proposes which learnings to persist as which artifact type. Companion to /skillify (which transforms external concepts into skills) — this one captures the internal learnings of a session into the right durable form.

When to use this skill

Use save when:

• The user explicitly says /save, "memorize this", "capture what we learned"
• A long session has produced multiple reusable patterns and the user wants to harvest them before context is lost
• After a productive back-and-forth that surfaced a new convention, anti-pattern, or tool usage worth keeping

Don't use save when:

• The session was a single one-off task with no transferable learning
• Auto memory will already capture it implicitly (the system writes some memories automatically — don't duplicate)
• The user is asking to save a specific file (that's just Write, not this skill)

The 5 categories of artifact

Each captured learning belongs in exactly one home. Pick the right home — wrong home = lost or duplicated.

If the learning fits none of the above — drop it. Not everything is worth persisting.

The procedure

Track each step with TodoWrite from the start.

Step 1 — Review the session

Re-read the conversation context (the user's recent messages + your own actions). Identify the non-obvious things — patterns the user explicitly validated, mistakes you made + corrections, recurring conventions, tools/URLs the user shared, decisions with rationale.

Skip everything that:

• Is already in MEMORY.md (run Read ~/.claude/projects/<project>/memory/MEMORY.md first to know what's already there)
• Is in CLAUDE.md (Read CLAUDE.md)
• Is trivial / one-off / "what to do next" task state

Step 2 — Categorize each candidate

For each learning, propose its home with the table above. Use this template :

1. <one-sentence description of the learning>
   → <Memory | Agent | Skill | Hook | CLAUDE.md>
   → Path: <proposed path>
   → Why this home: <one sentence>

If you're hesitating between two homes, default to Memory (it's the most reversible).

If two learnings are about the same topic, merge them. Don't fragment.

Step 3 — Confirm with user

Present the categorized list via AskUserQuestion. Let the user :

• ✅ Approve all
• ✏️ Edit specific items (re-categorize, rename, drop)
• ❌ Drop some

Don't write anything yet. Wait for explicit OK.

Step 4 — Write artifacts

For each approved learning, create the file with proper structure :

Memory note : YAML frontmatter (name, description, type ∈ user/feedback/project/reference) + body. For feedback/project types, include Why: and How to apply: lines.

Sub-agent : YAML frontmatter (name, description ≤ 200 chars, tools, optional model/permissionMode) + body system prompt.

Skill : YAML frontmatter (name, description includes trigger phrases, argument-hint, allowed-tools) + body with When-to-use / Procedure / Examples / Anti-patterns. Cap at ~250 lines. If extracting from external source, prefer /skillify instead.

Hook : add to .claude/settings.json hooks: array with proper matcher + handler. Script in .claude/hooks/<name>.sh made executable.

CLAUDE.md addition : append a clearly-scoped section. Don't dump huge blocks. Project-specific only — don't duplicate the global file.

Step 5 — Update indices & finalize

• For memories: append a line to MEMORY.md index in the format - [Title](file.md) — one-line hook
• Verify nothing leaked into the wrong scope (user-global vs project-local).
• Propose a single commit if applicable: chore: capture session learnings (memories + skills + ...).

Examples

Mid-session save (recommended cadence): after a productive 1h+ session that introduced a new convention. Run /save to harvest before context closes.

Post-test save: after testing a third-party tool (like Ruflo), /save is the natural follow-up to /skillify — /skillify extracts the idea, /save captures the meta-learnings (what we learned about the testing process itself, the verdict pattern, the watchlist).

End-of-day save: at the end of a long working day, /save to consolidate before tomorrow.

Anti-patterns

• ❌ Saving everything. Apply quality bar — if a future session won't gain from this, drop it.
• ❌ Putting a workflow into Memory instead of a Skill. Workflows belong to Skills (or Agents).
• ❌ Putting a one-off project decision into the user-global ~/.claude/CLAUDE.md. Project-specific stays project-local.
• ❌ Skipping confirmation. Always present the categorized list before writing.
• ❌ Re-saving things already covered. Read MEMORY.md and existing skills/CLAUDE.md first to avoid duplication.
• ❌ Modifying the user-global ~/.claude/CLAUDE.md without explicit text confirmation. Some hooks block self-modification — fall back to printing the suggested change.

Companion skills

• /skillify — when the learning is a concept observed in an external tool, transform it into a skill (Step 4 then becomes "write a skill file via the skillify procedure").
• (future) — /load could be its mirror : load relevant memory/skills at session start based on the current task.