loop-designlisted
Install: claude install-skill mwgrant21/agent-improvement
# loop-design
Gate for creating or modifying recurring loops. A loop that does not pass this
gate does not get scheduled.
## Before designing
1. Read `~/agent-improvement/loops/README.md` (the convention: state schema,
run-log schema, graduation rules).
2. Read `~/agent-improvement/domains/loop-design.md` and honor every lesson.
3. If modifying an existing loop, read its `LOOP.md`, `STATE.md`, and the tail
of its `runs.jsonl` first.
## Required declarations (creation)
Refuse to scaffold until ALL are explicit:
- **Name** - kebab-case, becomes `loops/<name>/`.
- **Trigger** - SessionStart hook / cron / scheduled agent / manual /loop.
- **Level** - ALWAYS 1 at creation. Do not accept a higher starting level.
- **Budget** - `soft` or a token cap.
- **Attempt cap** - default 3.
- **Sources/scopes** - what it reads; external connectors start read-only.
- **L1 boundary** - what "report-only" means for this loop, stated in LOOP.md.
- **Verifier plan** - required before any future L2 promotion (who verifies,
default stance REJECT). Recorded in LOOP.md even while the loop is L1.
- **Kill switch acknowledgment** - the runner must check `paused` first.
## Scaffolding (creation)
1. Create `loops/<name>/STATE.md` from `loops/_templates/STATE.md.template`
(replace `{{LOOP_NAME}}`).
2. Create `loops/<name>/LOOP.md` - the run protocol: sources, per-run steps,
state-update rules, run-log line, critique, retrospective rule, L1 boundary.
3. Create empty `loops/<name>/runs.