Jay Margalus
◆ Agents · Project 3 of 5 · 2 to 3 hours

It brings it up first

Until now you asked and it answered. This flips it. A run happens every morning without you, writes a short brief, and puts it where you'll see it. You also write down what's allowed to interrupt you.

Show

The top of Jay's morning prep page: Sunday, November 1. Nothing on the calendar today. Eight things waiting on a reply, and a Monday that starts tight.
6:30 this morning, written by the agent on my Mac mini from my calendar and mail. I read it over coffee. I did not ask for it.

Two jobs run for me every morning. At 5:45 one gathers my calendar, the last day of mail headers, and what changed overnight, writes a one-page brief, and sends it to the e-ink tablet in the kitchen. At 6:30 another writes the page above and ends with three offers of what it could do today. Both are listed in one manifest, the only place anything on that machine may be scheduled.

Why this is the real shift. A chat box waits for you. A scheduled agent doesn't, and the question changes from “what should I ask” to “what may it tell me, and when.” My answer is a file called ATTENTION.md: what interrupts, what batches, what it may do without asking. It's the most important file I have.

Do 0 of 7

  1. Run the agent without a conversation. In the practice folder:

    claude -p "Run the brief skill and write brief.md" --allowedTools "Read,Write,Glob,Grep"

    Headless mode (docs): one prompt in, one result out. If a job can be a sentence, it can be scheduled.

  2. Feed it something real, read-only. Export your calendar for the week (Google and Outlook export .ics; Apple Calendar reads with a one-line AppleScript). Put it in inbox/. Don't give it your mail password. Headers or an export are enough.

  3. Write the morning skill. A SKILL.md called morning: read inbox/, write briefs/YYYY-MM-DD.md with today's schedule, what's waiting on a reply, one thing to prepare. Under 200 words, with sources. Run it headless once by hand.

  4. Schedule it. Mac: a launchd plist in ~/Library/LaunchAgents at 6:00 (Apple's guide, or ask Claude Code to write it and tell you how to load it). Linux or a Pi: crontab -e, after backing the crontab up. A plugged-in laptop works for a week. Then you'll want a machine that's always on.

  5. Put it where you'll read it. A brief in a folder is useless. Open it in your notes app automatically, email it to yourself (agent drafts, script sends), or push it to a tablet. Pick the one you'll see before 8am.

  6. Start the manifest. One file, manifest.json or a plain table: job name, schedule, what it reads, what it writes, who to blame. Every job goes in before it's scheduled. In a month you won't remember what runs.

  7. Write ATTENTION.md. Three short sections. Channels, most expensive to cheapest (a text, an email, the brief). What's auto-approved (read the folder, write briefs) and what's gated (send, delete, spend). When unsure: batch it into tomorrow's brief. Add to CLAUDE.md: Read ATTENTION.md before contacting me.

Ticks are saved in this browser.

Check

  • A brief appeared this morning that you didn't trigger, and it named something real from your calendar.
  • The job is in the manifest with schedule, inputs, and outputs.
  • ATTENTION.md answers: what may interrupt me, what waits for the brief, what needs my yes.
  • Kill test: turn the job off for a day using only the manifest. Over two minutes means the manifest isn't the source of truth yet.

In your course

Assign

One scheduled job on real inputs from a real person outside the class, running five working days, with the manifest and ATTENTION.md turned in.

Grade

Five days of output next to what the process owner did with it. One honest “it got Tuesday wrong and here's why” beats five clean briefs.

Ask weekly

Two numbers out loud: conversations with real people, and touch points running on real inputs. Zero is fine to say and not fine to hide.

The whole term's plan is on the faculty guide.