Apprentice — Watch Me Once. Do It Forever.

The entire history of computing: you describe what you want → the computer executes.
Apprentice flips it: you do what you want → the agent watches → it becomes a permanent skill.

This is Programming by Demonstration — a 30-year holy grail of human-computer interaction research. Every attempt failed because it required constrained environments or rigid formal specifications. LLM agents make it possible for the first time:

• - Watch what you actually do (not what you say you do)
• Understand your intent, not just your actions
• Generalize across contexts — knowing what's a variable vs. what's a constant
• Turn it into a repeatable workflow your agent can run, refine, and chain forever

External Endpoints

Apprentice records locally. All synthesis happens via your already-running LLM session. No external APIs.

Security & Privacy

• - Zero external calls. Observation logs, workflow files, and all synthesis happen locally.
• No credentials accessed. Apprentice observes what you tell it — it doesn't read system files or intercept your input at an OS level.
• You review before saving. After observation, you see the synthesized workflow and approve it before it's saved as a permanent skill.
• Sandboxed execution. Generated workflow run scripts are plain bash with full security manifest headers.

Trust Statement: Apprentice writes observation logs and workflow SKILL.md files locally. The only thing it sends to the LLM is your described observation + synthesis request, using your existing session. Nothing leaves your machine.

Model Invocation Note

Apprentice is always user-initiated. It activates on "watch me" or equivalent phrases, and pauses until you say "done" or "stop watching." Synthesis runs once per observation session, with your review before saving. Workflows execute only when explicitly invoked.

How It Works

Step 1 — Start Observation

"Watch me"
"Apprentice mode on"
"Learn this"
"I'm going to show you how I do X"

Step 2 — Do the Task Naturally

• - Commands you run
• Files you create, edit, or delete
• Decisions you make and why
• The order of steps
• What changes between runs (variables) vs. what stays constant

You don't need to be formal. "Now I update the README with the project name" is enough.

Step 3 — Stop Observation

Step 4 — Review the Synthesis

• - What it understood as the workflow name and purpose
• The steps it identified, in order
• The variables it detected (things that change each run)
• The constants (things that stay the same)
• The generated workflow as a readable, editable description

You approve, edit, or reject.

Step 5 — It's a Permanent Skill

"Run new-project-setup"
"Do the weekly-report workflow with topic=AI agents"
"What workflows have you learned?"
"Replay morning-routine"

Example: Learning a Workflow

User:
CODEBLOCK3

[User does the following, narrating out loud]:
CODEBLOCK4

User:
CODEBLOCK5

Apprentice responds:
CODEBLOCK6

Next time:

"New project setup for analytics-redesign"
→ Agent runs all 5 steps automatically with PROJECT_NAME=analytics-redesign

Trigger Phrases

Start observation:
CODEBLOCK8

Stop observation:
CODEBLOCK9

Run a workflow:
CODEBLOCK10

Manage workflows:

"What have you learned?"
"Show me my workflows"
"Delete [workflow-name]"
"Edit [workflow-name]"
"When did you learn [workflow-name]?"

What Makes a Good Observation

The more context you give, the better Apprentice understands:

✅ Talk out loud — "Now I'm going to update the config with the new API key"
✅ Explain why — "I always do this before deploying because staging needs different env vars"
✅ Name the variables — "The PROJECT_NAME here would change each time"
✅ Mark the end — "And that's the whole workflow, every time"

❌ Silent actions — Apprentice can only learn what it can observe through your conversation
❌ GUI-only tasks — Apprentice works with what you describe; it doesn't watch your screen

Workflow Files

Each learned workflow lives in apprentice/workflows/<name>/:

CODEBLOCK12

The generated SKILL.md is a full, valid OpenClaw skill. This means:

• - Other skills can call it
• You can edit it manually to refine
• You can publish it to ClawHub to share with others

Chaining Learned Workflows

Once you have multiple workflows, Apprentice can chain them:

CODEBLOCK13

Workflow Library

After a few weeks of use, your workflow library becomes a personal operating system — a library of you. Things like:

• - morning-routine — the first 15 minutes of your day
• INLINECODE4 — how you start every project
• INLINECODE5 — how you compile and send the Friday summary
• INLINECODE6 — every step you take when a new client joins
• INLINECODE7 — your exact deployment sequence
• INLINECODE8 — how you prepare before reviewing a PR

No two users' libraries will ever be the same. This is your agent, shaped by what only you do.

File Structure

CODEBLOCK14

Philosophy

Every tool ever built asks: What do you want?

Apprentice asks: Can I watch?

The difference is everything. When you describe what you want, you lose nuance — the order matters, the edge cases matter, the "I always do this first" matters. When Apprentice watches, it captures all of it, exactly as you actually do it.

Your agent doesn't get smarter by being trained on more data. It gets smarter by watching you.