Day Zero · Session 04 · Dates TBD
A support agent and a workflow that runs while you sleep
Ship AIPhoenix, ArizonaPress → to begin
Until tonight, everything waited for you to show up.
The skills, the site, the drafts — all of it starts when you open a laptop. Tonight two things stop waiting.
Four sessions, one business — yours. Every session is free and standalone, each one leaves a folder the next one reads, and catching up late is explicitly allowed.
Act one
Two different shapes with two different failure modes. Naming which one you're building decides how you test it.
| Agent | Workflow | |
|---|---|---|
| Trigger | Someone asks | The clock |
| Shape | A conversation, one turn at a time | Fixed steps, start to finish |
| Fails by | Answering confidently and wrong | Failing silently at three in the morning |
| You test it by | Asking it ten real questions | Running it by hand and reading the logs |
| Tonight | Support, on your own site | Drafts, into a review queue |
Neither of these is smarter than the other. One is a front door and one is a factory.
The handoff
git fetch && git switch night-shift
app/api/chat/route.ts — newworkflows/weekly.ts — new3 packages added
Tonight extends the exact repo you shipped in session 03 — no fresh codebase, no second setup night. The branch brings an empty chat route and an empty workflow. You write what goes in them; the plumbing is already wired.
The branch is pre-tested. If yours doesn't start, we fix it in the room rather than at the end.
Act two
A chat box on your site that answers from your own content, and knows the difference between a question it can answer and one it shouldn't touch.
Answer, escalate, refuse — three legal outcomes and no fourth. An agent with no way to fail will invent one.
This file gets written before you deploy, not after the first screenshot of it going wrong.
Build it, then try to break it
/support-agent
answers from your pages, not model memory3 outcomes: answer · escalate · refusewrote 04-agents/guardrails.md
The prompt is built from the site you already shipped, so it cites your content instead of remembering something plausible about your industry. Then ten real questions — seven it answers, two it hands to a human, one it refuses.
If it answers one of the hand-off two, it isn't ready. That's the test, and it takes five minutes.
Act three
One workflow, running on a schedule, doing the part of your week that never needed you. Pick one tonight — the other one is a step in the guide.
Nothing publishes itself. The queue is the checkpoint from session 02, wearing a different hat — and the numbers path is the same loop with your report at the end of it.
Manual first, always
/scheduled-workflow
manual trigger first — run #1 in 4.2slogs read, one failure forcedcron written to vercel.json — Mondays 13:00 UTC
Run it by hand, read the logs, then turn the schedule on. Then break it on purpose and confirm you hear about it — a cron job that fails silently is worse than no cron job, because you'll trust it for a month.
The agent is everyone's. The two middle rows are a pick-one — you build whichever you chose tonight, and the other is a step in the guide and an evening this week. Four sessions ago this was a blank box in a chat window, and the bottom row is zero on purpose: it's the one you keep.
Feel the fear and do it anyways.
Next
Day Zero got it running. Zero to Launch is the one that puts customers in front of it — funnel, positioning, distribution, and a hackathon at the end where you ship it in public. Tonight's deck, guide and skill files are on the session page.
shipai.club · Free, always