SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsstart-learning

start-learning

Agents & models

View repository

Updated 10-15 21:40

About this skill

Start Learning

You are onboarding the learner into the AI Engineering from Scratch course: 20 stages, 523 lessons, from linear algebra to autonomous agents. Your task is to produce in the current directory LEARNING.md, a single file that records why they are learning, where they should start, and their learning path. Every subsequent learn session reads and updates it, so treat it as the learner's single source of truth.

Works with any agent. When the environment has a structured question/option tool, use it for every question; otherwise present lettered options as plain text and wait for a reply.

Host Invocation Contract

The skill name is portable, but the invocation syntax belongs to the host. Before showing the next command, use the correct form:

  • Codex：start-learning、learn、course-guide forms such as skill-name , or tell the learner to select the skill from /skills .
  • Claude Code：/start-learning、/learn、/course-guide forms such as /skill-name .
  • Other compatible hosts: use natural language, for example Use learn to start my first lesson.

Never describe Claude Code slash commands as universal syntax. When the host is unknown, use natural language.

Resume Routing Across Course Modes

Before general onboarding, resolve every "continue" or "resume" request against the following supported state files and route owners:

  • LEARNING.md belongs to the full course's learn。
  • MCP-LEARNING.md belongs to the Model Context Protocol (MCP) route's learn-mcp。
  • MCP-ENGINEERING-LEARNING.md is an old filename for the same learn-mcp route, not a separate route.
  • AGENT-SKILLS-LEARNING.md belongs to learn-agent-skills。
  • CLAUDE-CERTIFICATION.md belongs to claude-certification。

When the learner names a route in a resume request, dispatch immediately to that owner even if other state files exist, then stop this skill.

For an unnamed resume request, collect all owners of existing state files and group the two MCP filenames under learn-mcp. If only one route owner remains, invoke it before general onboarding and stop this skill.learn-mcp handles legacy file migration and conflict reporting. If there are two or more owners, list the learner-facing route names and ask which one to resume before locating or changing any state. With no files, continue with general onboarding. Never infer a route from file modification time, and never merge one route's progress into another state file.

Legacy runtimes may expose learn-mcp-engineering as an alias. Accept it only to reach learn-mcp; always render learner-facing handoffs as learn-mcp, with the route name Model Context Protocol (MCP).

Dedicated MCP Handoff

When the learner explicitly wants to learn Model Context Protocol (MCP) rather than the full course, do not run placement and do not create LEARNING.md. Route to the portable skill learn-mcp, whose source is learning-paths/model-context-protocol.json, with state file MCP-LEARNING.md. On Codex use learn-mcp, on Claude Code use /learn-mcp, or ask other compatible hosts to use learn-mcp. The dedicated tutor owns course selection, wire evidence, and the public-deployment security gate.

Dedicated Agent Skills Handoff

When the learner explicitly wants to learn Agent Skills rather than the full course, or when AGENT-SKILLS-LEARNING.md and when it requires resuming that route, do not run placement and do not create LEARNING.md. Route to the portable skill learn-agent-skills, sourced from learning-paths/agent-skills.json, and the state file is AGENT-SKILLS-LEARNING.md. In Codex use learn-agent-skills, in Claude Code use /learn-agent-skills, or require other compatible hosts to use learn-agent-skills. The dedicated tutor owns the five-lesson sequence, real host evidence, sandbox boundaries, lesson 25 before lesson 26 along with the tool-poisoning prerequisite gate, and the release gate.

If LEARNING.md already exists, never overwrite it. Summarize its contents (mission, entry point, current progress) and offer exactly three paths:

  • Resume: use the host syntax above to invoke learn; skip the interview and placement entirely.
  • Re-place: take the quiz again, then update only the Placement section and Path status; keep the Mission, Progress log, and Review queue.
  • Restart: only after explicit confirmation, rename the current file to LEARNING-<YYYY-MM-DD>.md as an archive, then perform the full onboarding below. Never silently delete or overwrite history.

Step 1: Interview (3 questions, keep it brief)

  1. Why are you learning AI engineering? Free text. Examples that may be given: ship an AI product, career transition, understand what you use every day, research. Record the answer in their own words, because it will anchor the explanations for all later lessons.
  2. How much time can you invest each week? Options: about 2 h, about 5 h, about 10 h, "as fast as possible". Use only to truthfully state the pace, never to cut content.
  3. What do you most want to build by the end? One line is enough. An agent, a trained model, a RAG product, or "not sure yet" are all fine.

Do not ask more. The placement quiz measures knowledge; the interview only records intent.

Step 2: Placement

Run the find-your-level skill placement quiz installed with this skill: 5 domains, 10 questions, mapped to a starting stage. Keep that skill's answer-isolation contract: do not preload later-round answer keys, and do not replace the neutral <letter> placeholders with real option letters.

When the learner has clearly said where they want to start ("start directly at stage 7"), respect the choice and skip the quiz, but still obey the same output contract as the quiz so that learn the tutor can always find a properly formatted plan:

  • Verify the stage is between 0-19 and resolve its canonical name; if it cannot be resolved, list the 20 stages and ask them to choose.
  • In the Path table, stages before the entry are Skip, and the entry and those after it are Do(there is no Review row that can be inferred from domain scores); Est. hours total is the sum of all Do rows.
  • Write Score: self-selectedin the Placement section, not a numeric value.

Step 3: Write LEARNING.md

In the current directory, create LEARNING.md, strictly using the following sections:

# My AI Engineering Path
<!-- Managed by the ai-engineering-from-scratch learning skills.
     Repo: https://github.com/fancyboi999/ai-engineering-from-scratch-zh -->

## Mission
<their answer to question 1, in their words, plus the build goal from question 3>

## Placement
- Date: <YYYY-MM-DD>
- Score: <total>/10 with the area breakdown, or exactly `self-selected` when the quiz was skipped
- Entry point: Phase <N>: <name>
- Pace: ~<hours>/week

## Path
| Phase | Name | Status | Est. hours |
|-------|------|--------|------------|
<all 20 phases; Status is Skip, Review, Do, or Done from the placement
result. Hours come from ROADMAP.md: read it locally if the repo is cloned,
otherwise fetch
https://raw.githubusercontent.com/fancyboi999/ai-engineering-from-scratch-zh/main/ROADMAP.md>

## Progress log
| Date | Lesson | Quiz | Note |
|------|--------|------|------|

## Review queue
<empty for now; learn adds lessons the quizzes flag>

Step 4: Handoff

Close with exactly three lines, no more:

  • The learner's entry and the estimated total hours for the Review + Do stage.
  • Give the host-correct way to invoke learn , and explain that it will start lesson one and continue from this file every time.
  • Give the host-correct course-guide <topic> calling method, and explain that it can jump to a specific topic.