The Night Shift

Two things that stop waiting for you: a support agent live on the site you shipped in session 03, and one workflow on a schedule. Guardrails get written before either goes live, and you break a run on purpose before you trust it.

Five-minute setup

Do this before the session starts, or in the first ten minutes while the framing happens. Nothing here is required to attend.

1. Have an agent installed

   Claude Code, the Codex extension, or whatever you already run. Every skill file below is plain markdown with frontmatter — it works in anything that reads a skills directory.

2. Get the skills

   Download this session's bundle and unzip it at the root of your project. It writes into .claude/skills/ and nothing else.

   unzip -o day-zero-04-agentic-workflows.zip -d .

3. Point it at a real product

   An idea in a notes app counts. The skills read your repo, your site and your README before asking you anything, so run them from inside the project rather than a blank folder.

4. Make somewhere to put the output

   The skills write markdown and don't care where it lands. One folder per session — 01-voice/ through 04-agents/ — at the root of your project. Make tonight's before you start rather than mid-run.

The steps

1. 01

   Pull the prepared branch and set your keys

   Same repo as session 03: git switch night-shift, which already has the chat route and the workflow wired up. You're writing what goes in them, not the plumbing. Keys go in .env.local and nowhere near a commit.

   Done looks like: The site runs locally on the branch and app/api/chat/route.ts is there — a wired route with nothing behind it yet.

2. 02

   Decide the rules before any prompt

   Never promise, never negotiate price, never send email, and "I don't know, here's a human" is always allowed. Decide them now and hand them to the skill in the next step — it writes 04-agents/guardrails.md as its first move, so anything you type into that file by hand gets overwritten. Deciding them after the first bad screenshot is how everyone learns them the expensive way.

   Done looks like: You can say all four rules out loud, and every one is a sentence you'd be happy for a customer to read.

3. 03

   Build the agent on your own content

   /support-agent

   The prompt gets built from the pages you shipped last session, so it answers with your content and cites the page it came from. Three outcomes only: answer, escalate, refuse.

   Done looks like: Every answer points at a page on your site. If it's talking about your industry in general, it's reaching for model memory.

4. 04

   Ask it ten real questions, two of which it must escalate

   Seven things a customer would genuinely ask, two it must escalate, and one it must refuse — a discount request and a delivery promise are the classics. Log the question, the outcome and the page it cited, and log what it said rather than what you hoped.

   Done looks like: It escalated both of the two. If it answered either one, fix the prompt and run the ten again.

5. 05

   Deploy it

   Onto the live URL from session 03. A chat agent that only works on localhost is a demo of a demo — and the first real question from a stranger teaches you more than the next hour of tuning.

   Done looks like: You can open the site on your phone and ask it something.

6. 06

   Run one workflow by hand

   /scheduled-workflow

   Pick one: weekly article drafts from your session-02 calendar into a review queue, or the numbers report. Trigger it manually, then read the logs line by line. Never auto-publish — the queue is the checkpoint.

   Done looks like: One complete run, and you can say what each step did without guessing.

7. 07

   Turn the schedule on, then break it on purpose

   Add the cron entry, then make a run fail — pull a key, point it at a file that isn't there — and confirm you actually hear about it. The alert has to live inside the workflow, not at the trigger: the trigger returns the moment the run is queued, so a catch there reports success on a run that died. A cron job that fails silently is worse than no cron job, because you'll trust it for a month.

   Done looks like: The failure alert reached you somewhere you'll see it. Then you put the key back and watched a clean run land.

What you leave with

04-agents/ — guardrails.md, the chat route live on your deployed site, a test log of ten questions with two escalations, and one workflow that has run manually, on a schedule, and failed loudly. Plus the optional /content-repurpose cut of your article in 02-routines/social/, if you took it.

If you get stuck

If the agent is confident and wrong, it's reaching for model memory instead of your pages. Tighten the prompt to answer only from retrieved content, and make "I don't know — here's a human" an explicitly valid answer. An agent allowed to say no is worth more than one that always has something to say.

Or ask in Discord — mentors and everyone else's work in progress are there between sessions.