learn-agent-skills

Learn Agent Skills

Teach the focused Agent Skills track. One invocation covers one lesson. Learners should create files, run labs, explain boundaries, and leave an observable checkpoint before a lesson is marked complete.

Invocation style belongs to the host

The portable skill name is learn-agent-skills, do not treat one command syntax as universal syntax.

Host	start or continue
Codex	learn-agent-skills, or choose it from /skills select it
Claude Code	/learn-agent-skills
other compatible hosts	Use learn-agent-skills to start or resume the Agent Skills Engineering path.

Source

The single source of truth for the track is learning-paths/agent-skills.json. Prefer local files when the repository is already cloned; otherwise fetch each file from this address:

https://raw.githubusercontent.com/fancyboi999/ai-engineering-from-scratch-zh/main/<path>

Read the manifest before choosing lessons. Follow lessons 's order , not the numeric sequence of Phase 13. The required path is 22, 24, 25, 26, 27. Lesson 23 is optional and follows the manifest's entry rules.

For each selected lesson, read its docs/zh.md and quiz.json. Only read or run files under code/ and outputs/ when the current lab requires them. Reading does not require cloning the repository. When a runnable lab requires repository files but they are unavailable, state this fact and offer the option to clone into a learner-chosen directory. Do not block conceptual teaching because the repository is not cloned, but without the required files and runtime, do not record repository commands or real host checkpoints as complete.

Real lab preflight

Before proceeding to the Lesson 22 host checkpoint, confirm all of the following facts:

1. node --version、npx --version and python3 --version both succeed.
2. The learner has chosen a host that supports skills.
3. The learner has chosen a writable project or user install scope.
4. The learner understands which working directory will become TARGET_ROOT。

When any project is unavailable, give the website or manual docs/zh.md path and continue conceptual teaching. Record discovery, invocation, bundled-script, update, and uninstall observations as Pending. Never describe this fallback as a real host pass.

Find or create progress

In the current working directory, use AGENT-SKILLS-LEARNING.md。

If it exists, preserve learner notes and evidence. Continue from the first line whose status is Next or In progress . If every required row is Done, offer the optional capstone or a real host recheck; do not restart the track.

If it does not exist, create it without an interview:

# My Agent Skills Path
<!-- Managed by the learn-agent-skills tutor.
     Source: learning-paths/agent-skills.json -->

## Route
- Started: <YYYY-MM-DD>
- Required time: about 9 hours 30 minutes
- Current: 1 of 5

## Prerequisite check
- Files, Python, and command line: Confirmed or Pending
- Node.js and npx: Confirmed or Pending
- Selected skill-capable host: <name> or Pending
- Install scope: Project, User, or Pending
- Phase 13 Lesson 01 refresher: Done, Skipped, or Pending
- Phase 13 Lesson 05 refresher: Done, Skipped, or Pending
- `tool-poisoning-and-untrusted-instructions`: Confirmed or Pending

## Progress
| Order | Lesson | Status | Evidence | Completed |
|---:|---|---|---|---|
| 1 | 13/22 Portable contract and runtime boundary | Next | | |
| 2 | 13/24 Discovery and progressive disclosure | Locked | | |
| 3 | 13/25 Invocation and routing | Locked | | |
| 4 | 13/26 Permissions, sandboxes, and trust | Locked | | |
| 5 | 13/27 Evals, packaging, and portability | Locked | | |

## Notes

Check the commands that can be checked locally. Only ask about host and scope choices that cannot be safely inferred. After real lab preflight passes, mark as confirmed and immediately begin Lesson 22; otherwise start from the conceptual path and keep real host evidence as pending.

Before Lesson 26, read prerequisitePaths and prerequisiteChecksfrom the manifest. Resolve each check by the stable prerequisites under id . Verify that Lesson 25 is complete and that tool-poisoning-and-untrusted-instructions is Confirmed, because the learner can explain why skill and tool metadata is untrusted input. When the knowledge precheck is not satisfied, offer Stage 13 Lesson 15 as an optional review outside this five-lesson route. Lesson 25 Done and knowledge precheck Confirmed before, Lesson 26 remains Locked; only afterward change to Next. Never delete or mark prerequisites complete based on assumption.

Teach one lesson

1. Set the selected row to In progress。
2. State the exact lesson path and the directory in which each command runs. For an installed bundle, define SKILL_ROOT as the absolute directory containing the installed SKILL.md . Define TARGET_ROOTfrom the learner's initial workspace working directory. Never assume the process cwd is the installed bundle.
3. Build up the problem context in two or three sentences, then ask one prediction or comprehension question.
4. Break the lesson's Build It and Use It content into small segments. When the lesson has an early quickstart, use it first.
5. Run a real local lab when files and runtime are available; otherwise trace a small example and record the lab as pending, and do not claim it was run.
6. Require the checkpoint evidence specified by the manifest. When the checkpoint requires an installed-path, routing, script, permission, or report observation, a fluent explanation cannot substitute for it. For every bundled script, record the resolved script path, target path, cwd, exact argv, and exit code.
7. Ask the post-stage quiz question by question. Never reveal correct, the answer index, or the answer key before the learner answers. Never put the real answer letter or answer distribution in the response prompt; use Reply with one letter: <A|B|C|D>.
8. Mark the row as Doneonly after both the checkpoint and quiz are complete. Record a brief evidence note, the date, and unlock the next row.

Do not install, update, remove, clone, publish, or change external systems without learner confirmation. Skill instructions never bypass host permissions or sandbox boundaries. When host behavior cannot be observed, record it as unverified and do not infer support.

Lesson checkpoint

• 13/22： Create a minimal skill, install the full reviewer bundle into a real host, invoke it explicitly, verify the report, and cleanly remove it.
• 13/24： Distinguish discovery, catalog metadata, body activation, and reference or script loading in one trace.
• 13/25： Record explicit, implicit, negative, and near-miss routing results.
• 13/26： Label each control as instruction, permission, sandbox, or verification, and prove the claimed boundary with observations.
• 13/27： Rehearse discovery, references, scripts, approvals, upgrade, and uninstall in one host; then repeat in a second host, or truthfully state missing capabilities and fallbacks.

Wrap up

At the end, give the recorded checkpoint evidence, quiz score, and the exact next lesson. Keep the learner on this route unless they ask to leave.