What if your scheduled AI jobs got a little better every day?
Most scheduled AI tasks run the same prompt forever. If the output is mediocre on day one, it is still mediocre on day thirty, and you pay the same for it. TinkerClaw’s jobs keep notes on their own runs and build on them. Here is how that works, and the one rule I added after it went wrong.
The board
On the day of this screenshot the board held 22 jobs, 18 of them on, with 2,711 runs in the history. A morning briefing, a security check, memory consolidation overnight, a scan of other forks for ideas worth borrowing, a model-ranking refresh. Each run records which model did it and how many tokens it took.
The loop
Each run ends with a review step, and the review has somewhere to go:
- Do the task. Write the briefing, run the scan, check the updates.
- Compare. Read the job’s lessons file and the last few runs. Is this better or worse? What did it miss?
- Write the lesson down. One dated entry in the job’s lessons file, which the next run reads before it starts.
- Propose, don’t rewrite. If the instructions themselves should change, the job writes the proposed change into its report. I approve it or not.
How to set one up
In OpenClaw and TinkerClaw, a cron job can carry an agentTurn payload that runs in an isolated session, with a fresh transcript every run. The trick is two extra lines in the brief:
Schedule: daily at 08:00
Brief:
1. Read lessons.md before you start.
2. Do the task (report, scan, check).
3. Compare with the last runs: what was missed?
4. Append one dated lesson to lessons.md.
5. If the brief itself should change, propose it in the report.Why a human approves the brief
The first version let jobs rewrite their own prompts. It worked, until it didn’t: a prompt that edits itself drifts, a little each night, toward whatever the last run happened to reward. Lessons are cheap to keep and easy to read back. A changed brief changes every future run, so it gets a second pair of eyes. The jobs still improve on their own; they just can’t redefine their job without asking.
What we saw in the first month
- Daily engagement report. Day 1 was a wall of numbers. By day 30 it highlighted trends, flagged anomalies and suggested actions.
- Fork scanner. Day 1 listed forks. By day 30 it sorted them by what they changed and recommended what to borrow.
- Wind-down summary. Day 1 was a diary. By day 30 it was a plan of fixes to the system.
Your scheduled jobs should know more tomorrow than they do today.
Everything on the site, by topic or date: Browse the Tinker Zone