Phase 2

Ripping Out the Belts

AI is no longer a tool you use - it's a component in the workflow.

Don't Write Prompts From Scratch

Ask AI to generate a prompt for you. Then iterate on that - much faster, much better.

You don't need to be an expert to get expert-level answers. The meta-skill: using AI to make AI better.

Caveat: AI can sound very confidently wrong. Always verify critical outputs.

✍️

Meta-prompting

  1. 1

    Think of a task where you've been unhappy with AI output

  2. 2

    Ask the AI to write you a detailed prompt:

    Write me a detailed prompt that will help me [do X] consistently. Include role, context, format, and constraints.
  3. 3

    Copy the generated prompt, paste it as a new message, and test it

  4. 4

    Compare: is the result better than what you got before?

Reusable Skills & Instructions

Doing the same thing multiple times? Ask AI to create a skill or instruction document so you can re-use it across sessions and tools.

Example: a “meeting summary” instruction that always outputs the same format.

✍️

Turn your prompt into a reusable skill

  1. 1

    Take the prompt you generated in the meta-prompting exercise

  2. 2

    Ask the AI to turn it into a reusable document:

    Turn this into a reusable instruction document I can paste at the start of any conversation. Add sections for: Role, Context, Task, Format, Constraints.
  3. 3

    Save the result - this is your first “skill”

How a Prompt Actually Works

Every “chat” with AI is actually a structured API call. Understanding this removes the magic and gives you control.

How Tool Calling Works

When AI “searches the web” or “reads a file,” there's an invisible orchestration layer between the chat and the action. The AI doesn't execute anything - it generates text that describes what it wants to do, and the system carries it out.

AI-to-AI Prompt Chain

This is Phase 2 thinking: AI is no longer a standalone tool - it's a workflow component. The output of one step becomes the input for the next.

Demo

Claude → Google AI Studio Prompt Chain

Use Claude to research a topic and generate a structured prompt. Copy that prompt into Google AI Studio. AI Studio builds a full interactive app from it. One AI preparing work for another.

Claude in Your Workflow (CoWork)

This is Phase 2 in action: AI isn't a standalone chat - it's embedded in your daily tools. CoWork (launched January 2026) is a persistent agent inside the Claude Desktop app. The key differentiator: shared context . Claude sees your Excel, PowerPoint, AND Gmail simultaneously.

Example: Claude reads your inbox, pulls a spreadsheet attachment, analyzes it, and drafts a reply with the findings. One workflow, three apps, zero copy-paste.

Demo

CoWork cross-app workflow

Claude reading an email attachment, analyzing it in Excel, and drafting a reply - all in one conversation.

CoWork is about ripping out the belts - instead of copy-pasting between apps, AI flows across your entire toolkit.

Dispatch

Send tasks from your phone to your desktop Claude. Scan a QR code, connected in 2 taps. Note: tasks are routed through Anthropic's cloud - not fully local processing.

Demo

Dispatch from phone

Scanning the QR code, sending a task from phone, watching Claude execute it on desktop.

When More Context Hurts

More context isn't always better. Long conversations dilute your original instructions. The AI doesn't ignore them on purpose - they just become a tiny fraction of what it's paying attention to.

Practical guidance: Start fresh conversations for new tasks. Re-state critical constraints. Keep system prompts concise. Exclude irrelevant context.

Death of the UI

Everyone sees data the way THEY want or need. Lower systems enforce integrity, authentication, authorization. The interface becomes a preference, not a constraint.

“Your UI is only a reference implementation. If your APIs and data structures are clean, customers will generate their own interfaces.”

Example: MES/SCADA/ERP enforce data integrity. AI becomes the personalization layer. An operator sees different data than a plant manager, without building separate dashboards.

A Warning from Samsung

When you start connecting systems, the stakes go up.

Samsung - Three Leaks in 20 Days

Lifting the ChatGPT ban without controls led to 3 source code leaks.

“Don't use it” and “use it freely” are both wrong. Phase 2 requires governance.

Ready to redesign the whole factory?

We ripped out the belts. The machines are rearranged. But we're still in the same building. What if we started from scratch?

Enter the new factory →