compoundlisted
Install: claude install-skill anttitevanlinna/agents-102
# /compound — corrections become load-bearing rules
This skill closes the compounding loop. Self-review captures corrections in free-form memory files; `/compound` promotes them into structured, schema-validated entries at `memory/compounded/` AND (when applicable) amends the `check_*.md` compendium that fires at generation time. **One edit, every future agent learns.**
Memory location: `/Users/anttitevanlinna/.claude/projects/-Users-anttitevanlinna-Projects-agents-102/memory/`
## When to invoke
- **End of session** — after Antti says "we're wrapping up." Capture the top 1–3 corrections/decisions from the session.
- **Immediately after a correction** — when Antti redirects an approach, and the correction is strong enough to carry forward.
- **On a pattern observation** — when the same kind of correction has occurred 2+ times across sessions.
Do NOT invoke for trivial edits ("fix a typo"), conversational back-and-forth, or ephemeral task state.
## The flow
### Step 1 — Classify
Ask (or infer from context):
1. **type:** `correction | pattern | decision | taste`
2. **surface:** `writing | sales_copy | student_facing | prompts | lectures | strategy_tie_in | pedagogy | content_creation | research | platform | strategy`
3. **severity:** `low | medium | high | critical`
- `high` if this correction has happened before (check `memory/compounded/` + `memory/feedback_*.md` + self-review correction log)
- `critical` if un-fixed, it ships to a buyer or student
### Step 2 —