Skip to main content
This is an internal AI-literacy training built for lab partners. It is aimed at engineering-minded colleagues who already use AI in their work, development, and research (vibe coding) but have not yet systematically understood the configuration and mechanics. The goal is not to teach you how to launch a tool, but how to use it correctly: understand how it works, treat configuration as a controllable engineering interface, and build the judgment to evaluate tools on your own.

​
The core claim

Most engineering-minded users are stuck on one gap: “using AI” is not the same as “using AI correctly”. Treating the chat box as a souped-up search engine and the coding agent as a clairvoyant autocomplete gives you fluctuating output quality, runaway context, silent errors born of blind trust, and a personalized experience nobody on the team can reproduce. The stance: modern AI tools are engineering systems that can be configured, constrained, and automated, and configuration itself is the leverage. The same model, with or without CLAUDE.md, Skills, Hooks, and correct privacy settings, produces output that differs by orders of magnitude. This training pulls that “configuration and mental model” layer into the open.
The primary reference is the Anthropic Claude family (Claude.ai, Claude Cowork, Claude Code), covered in depth; other tools (OpenAI, Google, GitHub Copilot, Cursor) appear as comparisons. Tool-configuration details were verified as of 2026-05; every fast-moving item is marked with its source and as-of date.

​
Four learning dimensions

The dimensions are not linear. Part III (judgment) runs through all of it: on a first pass, read I → II → III → IV, but in practice, every time you learn a setting, return to III and ask “is this actually useful to me?”

​
How to read: three routes

Different partners start from different places, so here are three reading routes.
Route A: by dimension, in order (default, for systematic learning)01-1 → 01-2 → … → Part II → Part III → Part IV. Build the full mental model first, then start configuring.
Route B: by tool (for “I only use one tool”)Read 02-1 (the layer model) first to build a general frame, then jump to 02-2 (if you use Claude) or the 02-6 comparison table to find your tool, then fill in 03-3 (security and privacy) and 04-2 (writing rule files).
Route C: by pain point (for those already using it, wanting to fix something specific)“Unstable output” → 01-4 + 04-1; “not sure whether to install a Skill” → 03-1 + 03-2; “worried about data leakage” → 03-3 + Appendix B; “want to automate a repetitive flow” → 04-4 + 04-6.

​
Content map

Part I Foundations

7 units: from “why correct usage” through the four engineering layers (prompt / context / workflow / harness) to the 2026 tool landscape.

Part II Configuration

6 units: the layer model, the Claude family, settings.json, claude.json, the four mechanisms, cross-tool comparison.

Part III Judgment

4 units: tool fit, beyond GitHub stars, security and supply-chain risk, personal benchmarks.

Part IV Customization

12 units: opens with the full .claude directory map, then CLAUDE.md, rules, commands, Skills, Subagents, Hooks, Plugins, MCP, CLI-first, and Agent Teams in depth.

Part V Case Studies

4 units: OpenClaw, Hermes-Agent, and NemoClaw dissected, plus a three-way route comparison.

Appendix

Settings cheat sheet, privacy and security checklist, glossary, further resources.

​
Part I: Foundations

​
Part II: Configuration

​
Part III: Judgment

​
Part IV: Customization, integration, and automation

​
Part V: Agentic case studies

​
Appendix

Don’t read everything before acting. After each Part II configuration unit, do it once on your own tool, then return to Part III and ask “is this actually useful to me?” The value of configuration only proves out inside your real workflow.