shadok-schedulerlisted
Install: claude install-skill shadok-ai/shadok-ai
# shadok-scheduler
Turn a natural-language request like *"watch the Google Ads budget every morning
and alert me when a cap is exceeded"* into a token-efficient scheduled job on
**this channel**.
Only works inside a shadok-ai channel — it needs `SHADOK_SESSION_ID` and
`SHADOK_PORT` in the env (set automatically). If they're missing, tell the user
this must be run from inside a shadok-ai session.
## The token-efficient pattern — DO THIS
Never schedule a bare LLM prompt for monitoring: it burns tokens every run even
when nothing changed. Instead, split detection (free) from wording (rare):
1. **Write a check script** that prints **nothing** when everything is fine, and a
short line **only** when there is something to report (a threshold crossed, an
error, a meaningful change). Prefer reusing existing skill scripts (e.g.
`~/.claude/skills/google-ads/scripts/…`) for the data. Save it somewhere
durable (a skill's `scripts/` dir). Test it once: it should be silent on a
normal day.
2. **Register a cron** with that check plus a short alert prompt. The server runs
the check on schedule (0 tokens); the agent is woken **only** when the check
prints something, and receives that output prepended to the prompt.
Result: quiet days cost nothing; the LLM is spent only to word a real alert.
For pure reporting where you WANT output every run (e.g. a daily figures digest),
omit the check — the prompt runs every time.
### What a check can see — secrets and cwd
A chec