How the start-learning Skill Onboards New Users in AI Engineering from Scratch

The start-learning skill conducts a deterministic four-step pipeline—resume check, three-question interview, placement quiz, and LEARNING.md generation—to create a personalized, persistent learning plan for every new user.

The start-learning skill acts as the canonical entry point for the AI Engineering from Scratch curriculum, implemented in the rohitg00/ai-engineering-from-scratch repository. Rather than dumping users into generic content, it captures intent, assesses prior knowledge, and produces a structured LEARNING.md file that persists across sessions. This approach ensures that whether you are a complete beginner or a practicing engineer, the curriculum adapts to your specific timeline and objectives.

Host-Agnostic Invocation Options

According to the Host invocation contract defined in skills/start-learning/SKILL.md, the skill supports three distinct syntaxes to accommodate different AI coding assistants.

Codex host:

start-learning

Claude Code host:

/start-learning

Free-form natural language (any host):

Use start-learning to begin the course.

This flexibility ensures that learners can invoke the onboarding flow regardless of which AI assistant or IDE they are using.

The Four-Step Onboarding Pipeline

The skill implements a strictly sequential workflow defined in skills/start-learning/SKILL.md. Each step produces deterministic outputs that feed into the next phase.

Step 1: Resume Routing and State Detection

Before initiating a fresh onboarding, the skill scans for existing state files to prevent accidental overwrites. It checks for four specific filenames: LEARNING.md, MCP-LEARNING.md, AGENT-SKILLS-LEARNING.md, and CLAUDE-CERTIFICATION.md.

If any file exists and the user requests a resume, control is immediately handed off to the owning skill (learn, learn-mcp, learn-agent-skills, or claude-certification). If LEARNING.md exists but the user wants to adjust their plan, the skill presents three options: Resume, Re-run placement, or Start over.

Step 2: The Learner Interview

When no valid resume path exists, the skill executes a three-question interview designed to capture qualitative intent:

• Why do you want to study AI engineering?
• How much time can you devote per week?
• What do you hope to build by the end?

These responses are stored verbatim and later injected into the Mission section of the generated learning plan, ensuring the curriculum remains anchored to the learner's specific goals.

Step 3: Placement Quiz via find-your-level

The skill delegates knowledge assessment to the companion find-your-level skill specified in skills/find-your-level/SKILL.md. This invocation presents a 10-question quiz covering five distinct knowledge areas.

The quiz result maps to an entry phase (numbered 0-19) and determines which sections of the 20-phase roadmap the learner can safely skip. This placement data feeds directly into the Placement section of the final document, including the week-by-week pace calculation.

Step 4: Generating the LEARNING.md File

Using the interview responses, placement outcome, and static roadmap data from ROADMAP.md, the skill writes a fully-structured LEARNING.md file to the project root. The file contains four critical sections:

• Mission – The learner's stated purpose and end-goal from the interview
• Placement – Date, score, entry phase, and calculated weekly pace
• Path – A markdown table of all 20 phases with status markers (Skip, Do, Review, Done) and estimated hours derived from ROADMAP.md
• Progress log and Review queue – Placeholder sections for future skill updates

The exact markdown template is defined in the Write LEARNING.md section of the skill specification.

Step 5: Hand-off to Curriculum

After file generation, the skill outputs three concise lines to the terminal:

1. A summary of the entry point and total estimated hours
2. The correct learn invocation syntax for the host (e.g., /learn for Claude Code)
3. The course-guide <topic> invocation for jumping to specific topics

This deterministic hand-off ensures the learner knows exactly how to begin their first study session.

Anatomy of the Generated Learning Plan

The LEARNING.md file produced by the skill follows a strict schema. Here is a truncated example showing the structure:

# My AI Engineering Path

<!-- Managed by the ai-engineering-from-scratch learning skills.
     Repo: https://github.com/rohitg00/ai-engineering-from-scratch -->

## Mission

I want to ship an AI product that can answer user questions, and I aim to build a personal assistant agent.

## Placement

- Date: 2026-09-02
- Score: 7/10
- Entry point: Phase 4: Linear Algebra
- Pace: ~5 h/week

## Path

| Phase | Name               | Status | Est. hours |
|-------|--------------------|--------|------------|
| 0     | Foundations        | Skip   | 4          |
| 1     | Linear Algebra     | Do     | 6          |
| …     | …                  | …      | …          |

The Path table references the canonical roadmap defined in ROADMAP.md, ensuring that hour estimates remain consistent with the curriculum's published standards.

Resuming Existing Progress

If LEARNING.md already exists when the skill is invoked, the onboarding flow enters a protection mode. Instead of overwriting the file, the skill reads the existing state and offers three explicit options:

• Resume – Continue from the last recorded phase
• Re-run placement – Retake the 10-question quiz to adjust the entry point
• Start over – Delete the existing plan and begin the interview fresh

This safety mechanism prevents learners from losing their progress history or personalized mission statements.

Summary

• The start-learning skill supports three invocation syntaxes (Codex, Claude Code, and natural language) for maximum compatibility.
• It checks for four state files (LEARNING.md, MCP-LEARNING.md, AGENT-SKILLS-LEARNING.md, CLAUDE-CERTIFICATION.md) to enable seamless resume routing.
• New users complete a three-question interview and a 10-question placement quiz before the system generates their plan.
• The output is a structured LEARNING.md file containing Mission, Placement, and Path sections that persist across sessions.
• The skill sources phase data and hour estimates from ROADMAP.md and delegates quiz logic to skills/find-your-level/SKILL.md.

Frequently Asked Questions

What happens if I already have a LEARNING.md file?

If LEARNING.md exists, the skill detects it immediately and presents three options: Resume to continue your current path, Re-run placement to retake the knowledge quiz and potentially adjust your entry phase, or Start over to delete the existing file and begin fresh. This prevents accidental data loss.

How does the skill determine my starting phase?

The skill invokes the find-your-level skill, which administers a 10-question quiz across five knowledge areas. Your score maps to one of 20 phases (0-19) defined in ROADMAP.md. The result determines which phases are marked as Skip versus Do in your generated LEARNING.md file.

Can I use the start-learning skill with any AI coding assistant?

Yes. According to the Host invocation contract in skills/start-learning/SKILL.md, the skill recognizes three syntaxes: bare commands for Codex (start-learning), slash commands for Claude Code (/start-learning), and natural language prompts for generic hosts.

Where does the skill store my interview answers?

The skill writes your responses to the Mission section of LEARNING.md verbatim. This ensures that throughout the 20-phase curriculum, your original goals and time constraints remain visible at the top of your learning plan, helping the system prioritize relevant topics.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}