Skip to main content
The CrossoverStage 3 of 3: The Proof

Walk away. Come back. Work is done. This is not a tool you operate. This is capacity you deploy.

5.7

The Autonomous Session

Estimated time: 30 min (15 away from keyboard)Estimated cost: ~$0.30Tool: Claude Code (Sonnet)
After this drill, you can:

After this drill, you have walked away from the keyboard while Claude Code worked, returned to find the work done, and understood in your bones: this is not a tool I operate. This is capacity I deploy.

Why this matters

You have been building toward this. Drill 5.3 gave you project memory. Claude Code already knows your project. Drill 5.4 gave you slash commands. Claude Code already knows your workflows. Now you are going to set Claude Code a 15-minute task, step away from your computer, and let it work. When you come back, you will find committed code, a log of what was done, and changes in your browser that were not there when you left. That is The Crossover, third stage. The proof. Before this moment, every AI interaction required your presence. You prompted. You reviewed. You iterated. You were the bottleneck. After this moment, you understand at a physical level that AI can execute while you are absent. The relationship has changed. You are not the operator anymore. You are the architect who defined what "done" means and trusted the system to get there. This is not magic. It is not automation for its own sake. It is what happens when you invest properly in context (CLAUDE.md), in workflow definitions (slash commands), and in a clear success criteria (what "done" looks like). The quality of the autonomous session is entirely a function of the quality of what you built before it. That is the lesson. Before you walk away — your work is safe. Claude Code operates on a feature branch, not main. You cannot accidentally deploy broken code to production. If anything goes wrong, you have three immediate recovery tools: git status — see exactly what changed git diff — read the exact line-by-line changes git reset --hard — undo everything Claude did, instantly The worst case takes 30 seconds to reverse. The best case is the course's defining moment. Walk away.

How to do it

  1. 1

    Choose a 15-minute task with a clear completion condition

    The task must have a definition of done that Claude Code can verify itself. Good examples: 'Add a search bar to the projects list that filters by title', 'Add character count display to all text fields', 'Add keyboard shortcut indicators to all buttons'. Bad examples: 'Improve the UX', 'Make it look better'.

  2. 2

    Use the MCPP-lite framework to write the task brief

    WHAT / CONTEXT / DON'T / GO / REPORT. Use the template below. This is the prompt you give Claude Code before you walk away.

  3. 3

    Start Claude Code, give it the task brief, and step away

    Literally stand up and walk away from your computer for 10–15 minutes. Get coffee. Take a walk. Let it work.

  4. 4

    Return and review the work

    Read the Claude Code report. Check the commits. Open the browser. Does it match what you asked for? What would you change in the brief to get a better result?

The prompt

PROMPT — The Autonomous Session Brief (MCPP-lite format)Model: Claude Code (Sonnet)Est. cost: ~$0.30
WHAT: [Describe the outcome, not the steps. "The app has a working search bar on the projects page that filters by title as you type."]

CONTEXT: [What Claude Code needs to know beyond CLAUDE.md. "The projects list is in src/components/ProjectsList.tsx. Search should filter the array client-side (no API call needed). Use existing Tailwind classes from CLAUDE.md."]

DON'T: [Hard constraints. "Do not change the overall layout. Do not add new dependencies. Do not modify the project card component."]

GO: Execute this autonomously. Do not ask for confirmation at any step. Make all necessary decisions using your best judgment and the project conventions in CLAUDE.md.

REPORT: When complete, tell me: (1) what files you changed and why, (2) what decisions you made that weren't specified, (3) anything that needs my attention or review, (4) how to verify the result.

Success criteria

  • You wrote a complete task brief using the MCPP-lite format
  • You stepped away and let Claude Code run without interruption
  • The task completed and you reviewed the commits and report
  • You can articulate: the quality of autonomous work is a function of brief quality
  • Something about how you relate to AI has shifted

Common mistakes

Being too afraid to walk away — worrying autonomous execution will break something

You're on a feature branch. `git reset --hard` is your undo button. The risk is lower than it feels. Claude Code cannot push to main, cannot delete committed history, and cannot deploy anything. The worst case is fixed in 30 seconds. The only way to experience The Crossover is to not be there when it happens.

Watching Claude Code the entire time instead of stepping away

The point of this drill is the absence. You need to actually not be there. Watching is fine for other drills — this drill requires you to walk away and come back.

Writing a WHAT that describes steps instead of outcomes

"Add an onClick handler to the search button that calls filterProjects()" is a step. "The user can type in the search bar and the project list filters as they type" is an outcome. Claude Code figures out the steps.

Treating an imperfect result as a failure

An imperfect first autonomous session is a learning artifact. What part of the brief would you change? The answer to that question is more valuable than a perfect result.