AI Conductor Academy

Core: 0 of 6 done

Teach-from-Demo & Skill Capture About 30–40 min ← All modules

Elective 3

Teach-from-Demo & Skill Capture

Course outcome 6Focus: credentials and personal data

By the end of this module, you’ll be able to:

  1. Design a Teach-from-Demo runbook with roles, sequence, gates, and risk controls for capturing a skill from a demo.
  2. Write a teach brief that says when to teach from a demo and when to invent, with sensitive-info rules.
  3. Apply credential and personal-data controls so demos never leak secrets into what gets saved.

How you’ll show it: your teach brief and runbook (download). Course outcome 6 of 6

What good looks like

This is the bar each part needs to reach (Meets). You’ll get feedback on each part when you check your work.

  • Teach vs. invent: for both sides, when it applies (“when…”/“if…”) and what the agent does, in your own words.
  • Sensitive-info rules: at least 3 rules picked.
  • Credential and personal-data rules: separate rules for passwords/keys/tokens and for personal data like names and emails.
  • Pre-publish gate: who checks, when, and what’s blocked until it passes.
  • Roles, steps, risk: at least 2 roles, 3+ steps ending where redacted work is saved, and risk controls of 40+ characters.

How you pass: every part at Meets or better. A “this looks like a secret” reminder never blocks your download.

The automatic check looks for a few signals of these. It can’t judge quality, so aim for the description, not the keywords.

Plan how an agent learns a skill from a live demo, and what it does when there’s no demo to learn from, without ever capturing secrets or personal data.

Big question How do you teach a workflow that an AI agent can also execute and improve?

Note: Never paste real passwords, keys, or tokens here. If something looks like a secret, we’ll remind you, but it won’t block your download.

Good fit if: you record product demos or SME walkthroughs and want agents to turn them into reusable skills.

How this pattern works

Teach-from-Demo turns something a person shows once into a skill an agent can repeat.

  • Teach from the demo when a real demonstration exists: the agent captures the exact steps it saw.
  • Invent only when there’s no demo: the agent drafts a path from what it knows, and it’s flagged for human review.
  • Capture safely: demos happen in real accounts, so passwords, tokens, and personal data must never end up in what’s saved.
  • Why: a captured skill is reused many times, so one leaked secret or wrong invented step spreads fast.

Runbook anatomy

Every elective produces a runbook: a short document a team can follow without you in the room. It always has four parts, in this order.

  1. Roles: who does what. Include at least one agent and one human who approves.
  2. Sequence: the numbered steps from the trigger to where the finished work is saved.
  3. Gates: checkpoints that stop the work until someone approves. Each gate says who checks, when, and what it blocks.
  4. Risk controls: what could go wrong (sensitive data, blast radius, failure) and how you’ll prevent or contain it.

Tip: E1 has a full annotated example runbook if you want to see all four parts filled in. See the E1 example

Stuck? See sentence starters
  • Teach: “When a live demo of ___ is shown, capture ___.”
  • Invent: “If no demo exists for ___, synthesize ___ and flag it for ___ to review.”
  • Credentials: “Use demo-only accounts; replace any password or token with [REDACTED].”

Your project: teach brief + runbook

Step 1 of 3 Teach or invent

When to teach vs. when to invent

For each side, say when it applies and what the agent does.

Step 2 of 3 Protect sensitive info

Sensitive-info rules (pick at least 3)

0 of 5 checked (pick at least 3)

How will you handle passwords, API keys, or tokens? Use placeholders only. 15+ characters.
Keep this separate from credentials: names, emails, customer IDs, and what gets redacted. 15+ characters.
Pre-publish gate

Step 3 of 3 Write the runbook

Roles (at least 2)
Steps (at least 3)

Walk through start to finish, ending with redacted work saved to a file or artifact.

Name risks like access, blast radius, or failure, and how you’ll handle them. 40+ characters.

Tip: Feedback is automatic. It looks for specific details, like a number, a named owner, or a “when”. It’s a quick check, not a grade.

Download as

← Back to all modules