Paul TakisakiTHE AI FIELD GUIDE / 01

FIELD NOTE 10 / USEFUL UNATTENDED WORK

Automate the check before the job.

Move repeatable work off your keyboard without losing the ability to inspect it.

1,407summer sessions with no human promptSource & definition ↗

A recurring prompt needs a definition of success, a failure path, and a way to stop. Build those while you can still watch the job run.

“The rule for letting go: the job needs a deterministic gate it cannot talk its way past.”

Paul’s working notes · September 2026

What it looks like

THE FRICTION

“Every day, improve my website.”

A BETTER STARTING POINT

Every weekday: inspect the approved source. If something changed: prepare a draft. Check: source links resolve and required fields exist. Publish: only after human approval. Stop: repeated failures or the run budget.

Illustrative examples you can adapt.

Put it to work

  1. 01

    Make one manual run work.

    Write down the inputs, output, schedule and timezone, owner, and allowed actions. Save a good result. If the task is only moving known values between systems, a normal script may be simpler than an agent.

  2. 02

    Add the pass/fail check.

    Use deterministic checks where possible: a schema, a test suite, a row count, or a known file. Define a retry cap, run budget, and failure destination. Make repeated runs safe so a retry does not send the same message or create the same record twice.

  3. 03

    Choose the model deliberately, then observe.

    Use a capable model for the uncertain decision and consider a smaller model for a proven, bounded procedure. Compare quality on real examples before switching. Confirm the actual provider and model used by scheduled jobs. In Hermes, scheduled work has its own documented execution and model-selection behavior.

Give this to your AI

Adapt the brackets to your task.
Help me turn [manual workflow] into a proposed automation. Specify the trigger and timezone, inputs, exact output, allowed actions, deterministic checks, retry limit, cost or run limit, duplicate prevention, failure notification, owner, and stop control. Start with a dry run that cannot publish, send, delete, or modify production. If a simple script is enough, recommend it. Do not create the schedule until the dry-run result has been reviewed.
WHAT TO WATCH FOR

My prompt history shows prompts falling 68% from June to August while commits stayed high. I attribute part of that shift to unattended work. It is an observation from one workflow, not a controlled productivity measurement.

THE HABIT TO KEEP

A repeatable job deserves a repeatable check.

Sources & further reading

Personal counts are from Paul’s prompt-history analysis. Tool guidance was checked against these sources in September 2026. Confirm current behavior in your installed version.

A USEFUL NEXT READ

Give every agent a job description.