Skip to content

How the Hack Works

You have two hours to turn a real work problem into a working agent.

You will pick a scenario, pick an altitude, follow the building blocks, and demo what you built. Coaches will roam. Your team will make choices together. The goal is not polish. The goal is a working thing you understand well enough to keep using.

The flow

  1. Pick a scenario — start with a real Microsoft work problem.
  2. Pick your altitude — Base, Builder, or Advanced.
  3. Choose your data path — use your own data privately, or use the provided synthetic pack.
  4. Build the smallest working version — get to v0.1 fast.
  5. Test it on a real prompt — see where it breaks.
  6. Patch the spec — change the rule, sample, or boundary that caused the miss.
  7. Demo the behavior — show the before, the patch, and the after.

Ground rules

  • Nothing sends. Your agent drafts for human review. It does not send email, book meetings, or make commitments on your behalf.
  • Keep private data private. If you use your own Microsoft 365 data, keep it on your own screen.
  • Build small. A short, sharp spec usually beats a long one.
  • Use observed behavior. What you actually did last time matters more than what you wish you did.
  • Portability matters. Prefer files and formats you can carry across tools.

What "done" looks like

Done means you can run your agent on a fresh prompt and get useful output that reflects the scenario.

For Scenario 1, that means a portable twin spec — soul.md, voice.md, and revealed.md — plus an agent that uses it to draft, decide, or critique in a way that feels specific to the person it represents.

What you'll walk out with

You leave with two things:

  • A working artifact you can keep improving after the hack.
  • A transferable skill: how to compress messy work behavior into a small, testable agent spec.

See also: Run of Show · Facilitator guide