Conversion System

Turn repeated manual work into automation you can actually trust — and refuse to automate what shouldn't be. The registry is the single source of truth: ~/.claude/automations/registry.md.

Step 1 — CAPTURE

Any task done manually twice gets a one-line entry in the registry backlog. No judgment at capture time. Format: - [ ] <task, one line> — seen <n>x, ~<minutes> each

Step 2 — QUALIFY (3-question filter)

Ask all three. 3 yeses = automate. Any no = leave manual, note why.

1. Repeats same shape? Same steps each time, not bespoke judgment each run.
2. Inputs digital? Everything needed is reachable by tooling (API, file, MCP). ⚠️ claude.ai-connector MCPs (Linear, Granola, Slack…) are session-auth only — they do NOT work headless. If inputs need them, the automation is in-session, never a LaunchAgent.
3. "Done" definable in one sentence? Objective end-state a checker can verify.

Then apply the build bar (from manager skill): name the hours/week saved (with a number) OR the outward output shipped. Neither → don't build, even with 3 yeses.

Step 3 — SPEC (template, mandatory before build)

TASK:      <one sentence>
TRIGGER:   <manual /command | /loop in-session | LaunchAgent clock-based>
INPUT:     <exact sources; flag session-auth'd ones>
RULES:     <constraints, incl. voice/verify rules from memory>
STEPS:     <numbered, each verifiable>
OUTPUT:    <artifact produced — a DRAFT/file/report, never a sent thing>
CHECK:     <how the run proves itself correct; what auto-verify exists>
ESCALATE:  <when it must stop and ask>

Hard rule, non-negotiable, lives in every spec: Never send / spend / delete / promise ALONE. Any outward message, payment, destructive op, or commitment requires the human. Automation prepares; the human ships. (This encodes two standing rules: show-draft-before-send, and verify-before-asserting: data claims in outputs must be verified live by the run itself, never carried from notes.)

Step 4 — BUILD

• Trigger type per scheduler doctrine (manager skill): clock-based → LaunchAgent with token auth + log + dedup + 1 retry; in-session → /loop ONLY if it passes the 4-condition gate in ~/CLAUDE.md; else a plain /skill run on demand.
• Day-1 requirements: appears in aios-cron-view, probed by stack-health (if scheduled); registry row moves from backlog → active with spec link.

Step 5 — TRUST LADDER

Every automation starts at rung 1. Track current rung in the registry.

Demotion: any wrong output → drop one rung immediately and log the failure in the registry. Two failures at rung 1 → back to manual, entry closed with reason. An automation is never "done proving itself" — rung 3 is the ceiling, spot-checks never stop.

Registry format (~/.claude/automations/registry.md)

## Active
| name | spec | trigger | rung | last check | failures |

## Backlog (captured, unqualified)
- [ ] ...

## Rejected (failed qualify — with which question failed)
- ...