Toby Water 為 AI 智能體設定五項常駐工作，公開實際運作系統

Most people use their AI assistant like a search box with better manners: ask, get an answer, close the tab, explain everything again tomorrow. I run a one-person digital-goods store with an AI agent as my only operator, and the difference between "assistant" and "operator" turned out to be paperwork. Boring, specific, load-bearing paperwork.

The agent doesn't improvise my business. It executes five standing documents — job descriptions I wrote after the manual versions of these jobs started eating my week:

1. A money radar. Twice a day it scans a fixed pool of sources and answers one question: what can I act on today? Fixed format — signal, why it's hot, where the money is, do-today, shelf life — maximum five items, everything else archived silently. The format is the filter: a signal that can't name its "do today" doesn't ship.
2. An open-source scout. Every repo that crosses my feed gets four checks — is it real (stars vs. commits vs. releases), what's the actual license situation (repo, weights, trademark are three different questions), what does it truly cost to run, and which of four monetization shapes does it fit. Verdicts are Adopt / Watch / Kill, kept in an append-only ledger so we never re-evaluate a dead repo or forget why a live one is waiting.
3. An auto-delivery system. Every five minutes: scan the orders inbox (and the spam folder — ask me how I learned that one), match payment evidence against a product→files mapping on three keys (link ID, product, amount, with a 1% crypto-residue rule), deliver once, log it, and escalate anomalies to a human instead of guessing.
4. A content pipeline. Topic cards with a mandatory "what does this feed" field, a pre-publish checklist, and localization by adaptation notes — hook, register, scene swaps, channel tags, disclosure — never raw machine translation.
5. A sample standard. Three rules for anything free: it must be real (a complete piece of the actual product), complete (closes its own loop), and pass the pay-impulse test. This one exists because we shipped technically-working, commercially-dead video samples and killed the product line the same day. The standard also requires dissecting 2–3 samples from the head of the niche before producing ours.

None of this is a prompt. Prompts are what you type; these are what the agent is responsible for, including what it must never decide alone and when it must stay silent. A radar that reports nothing on a quiet day is a feature. An agent that always has something to say is a liability with a chat window.

The full auto-delivery rules document is free — it's linked below, complete, copy-ready, including the mapping and log templates. I picked that module as the sample because it passes our own standard: a stranger can implement it today and get a complete unit of value from it. If the free document is this thorough, you've learned something real about the paid ones.

Free module: https://tobywater.pages.dev/agent-ops-free-module.md

An honesty note, because this genre needs more of them: as of today (2026-10-23) the store this system runs is pre-revenue — the system is live, sales pending. I'm selling the documents because they're real and inspectable, not attaching an income screenshot, because there isn't one. When there is, the pack gets updated and Pro buyers get the new edition free.

The pack (five modules + blank templates; Pro adds the annotated real radar/scout issues with the reasoning left in) is at Tobywater Digital.

If you run a one-person thing: which job would you hand an agent first? I'm genuinely curious whether delivery beats radar for other people too.

Disclosure: this store and its content are operated with AI assistance — which, given the topic, will surprise no one.