The Last Mile, From Your Phone
Dispatch GUI work from anywhere — with a human gate where it counts.
TUTOR WITH THEFOCUS.AI
Copy this prompt into Claude, ChatGPT, or any external AI assistant. It points the assistant to the course instructions and links it to your student profile to track your progress and customize observations.
You are not enrolled yet. Enroll to generate a Student ID to track lesson completions and store learning notes.
Delegate GUI work from your phone
This gets truly interesting when combined with Dispatch. The workflow: you are on your phone, you send a task from the Claude app, and Dispatch spawns a Cowork session on your desktop at home — with your files, your skills, and your browser.
From your phone via Dispatch, send a task that requires GUI interaction. Cornwall Market: Sarah is at the second store location. From her phone, she sends: "Process the three invoice PDFs in my Downloads folder — enter them into the accounting system using our categorization rules."
Dispatch spawns a Cowork session on the office machine. It reads the first PDF — Chen's Produce, $1,240, fourteen line items. Applies the categorization skill. Opens the accounting system in Chrome. Enters each line item. Moves to the next PDF. Pacific Foods, $3,890 — splits items across departments, flags a $620 line item for review. Third PDF — Ridgeline Coffee, $2,290, all to account 5200.
Summary arrives on Telegram: "3 invoices processed, total $7,420. One item flagged: Pacific Foods line item $620 (organic flour, case qty — verify before approving). All categorizations applied per skill rules."
Sarah reviews the flag from her phone, confirms it is correct, and goes back to work. The invoice processing that would have taken her 30-45 minutes happened while she was managing the other store.
This is the pattern that changes the economics of administrative work. Not because AI can think about invoices — it has been able to do that for years — but because it can now operate the applications that process them.
Everything you have built
Here is everything you have used across Chapters 01 through 08:
| Tool | Chapter | What It Does |
|---|---|---|
| Claude.ai | Ch 01 | Ask questions with context, discover the legibility gap |
| Voice, camera, handoff | Ch 02 | Capture knowledge by dictation and photo, work across devices |
| Artifacts | Ch 03 | Working web pages and tools, shared with a link |
| Projects + Skills | Ch 04 | Persistent knowledge, formalized into SKILL.md |
| Cowork | Ch 05 | Analyze real data, experience trust inversion |
| Connectors | Ch 05 | Hosted MCP — Drive, Gmail, Calendar in one click |
| Claude Code | Ch 06 | Terminal agent with filesystem access, skills operational |
| Remote Control | Ch 06 | Phone as window into local session |
| Cloud sessions | Ch 06 | Work on Anthropic’s infrastructure, handoff via /teleport |
| MCP connectors | Ch 07 | Live data from QuickBooks, databases, APIs |
| Custom MCP server | Ch 07 | Your own tools and data exposed to Claude |
| Routines | Ch 07 | Recurring cloud workflows, Monday morning reports |
| Channels | Ch 07 | Event-driven responses, invoice alerts |
| Chrome DevTools MCP | Ch 08 | Browser automation, supplier portals |
| Claude in Chrome + Computer Use | Ch 08 | GUI automation, invoice processing via screen |
| Dispatch | Ch 08 | Phone-to-desktop task delegation |
You have gone from L0 (asking Claude.ai a question and getting generic answers) to L4 (delegating invoice processing from your phone while Claude operates desktop applications using your categorization rules against live data on a schedule).
The next two chapters are for teams ready to build production systems. Chapter 09 introduces evals and habitats — persistent agents that learn from corrections. Chapter 10 puts it all together into a compound system where every piece multiplies with every other piece.
Check your understanding
Answer in your own words — write it down before opening the key. Your tutor grades against the criteria and generates fresh variants on retries.
Q1. Where must a human remain in the loop even in a fully automated pipeline, and what does a well-designed approval point look like (use the invoice flow as your example)?
Answer key — attempt every question first
Answer key
Q1
Model answer: Humans stay wherever actions are consequential and hard to reverse — money moving, orders placed, filings submitted. A good approval point arrives with the work already done and the evidence attached: ‘Invoice entered, $3,890, one line item flagged ($620, likely case quantity) — approve?’ One decision, full context, reversible until approved.
Pass criteria: boundary = irreversible/consequential actions; approval design = completed work + evidence + single decision
Next: Run Your First Eval