create-alert

Author a **scheduled alert**: a claude.ai cloud routine that evaluates a per-alert *trigger condition* on a schedule and drops a message into Slack when it fires. "Did X change / happen / cross a threshold" is a free-form predicate the routine's agent evaluates each run — not a fixed catalog. This skill's job is to grill the fuzzy ask into an unambiguous, testable spec, prove it with a live dry-run, and stand it up only after you sign off. **Create nothing until the final synthesized spec is explicitly signed off.** No routine, no `.ai/routines.md` entry, no `settings.json` write, and **no Slack message of any kind** — until the user has read the assembled spec (informed by the dry-run) and explicitly approved *that written spec*. Assembling first and confirming later defeats the skill: a vague trigger becomes a false-firing (or silently never-firing) alert, and the dry-run + sign-off is what catches it before it goes live. `--spec-only` stops after the spec (no create, no settings write, no message). ## Hardening block (baked into every routine) ```! D="${CLAUDE_PLUGIN_ROOT:+$CLAUDE_PLUGIN_ROOT/skills/create-alert}"; D="${D:-$HOME/.claude/skills/create-alert}"; cat "$D/HARDENING.md" 2>/dev/null || echo "(HARDENING.md not found — abort and report a broken install)" ``` The block above is inserted **verbatim** into every routine prompt this skill creates. It is what keeps an unattended routine — which fetches untrusted web content while holding the baked-in Slack destinati...