docs(direction): AI-drives-git reframe, lesson=theory, Claude Code as example

Lock the course direction before applying it across modules:
- AGENTS.md: the git reframe (concepts + "the AI drives git"; arithmetic-by-hand
  in M1-3, calculator/AI from M4 on; don't re-teach basics); Claude Code as the
  concrete worked example while staying model-agnostic in principle; lesson-vs-lab
  rule (lesson = theory only); no-slop voice rule (kill "prose" etc.).
- _TEMPLATE.md: Key concepts = theory with example output (never "run this");
  Hands-on lab is the only place things run, and from M4 on the learner directs
  the AI to do git/setup and verifies.

Foundation for the M1-6 fixes and the M7-27+capstone sweep.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TfzV5QvtPDz8LJS3Pu5VLT

_TEMPLATE.md

@@ -39,9 +39,10 @@ By the end of this module you can:
 
 ## Key concepts
 
-*The actual teaching content, in prose, with commands and snippets inline. This is the bulk of the
-module. No fixed length — go as deep as the topic needs and no further. Use subheadings freely.
-Reframe an ops instinct the reader already has wherever you can.*
+*The teaching content — **theory only**. Explain the concept and why it matters; reframe an ops
+instinct the reader already has. To show a command, show it **with example output** as illustration —
+do NOT tell the reader to run anything here (all hands-on is the lab, and the lesson must not
+duplicate it). No "prose"/slop words. No fixed length — go as deep as the topic needs, no further.*
 
 ---
 
@@ -55,9 +56,11 @@ differentiator; never skip it.*
 
 ## Hands-on lab
 
-*A practical exercise that uses AI **and** the tool together, run on the reader's own machine. This
-is a tools course — end at a keyboard, not a quiz. State the lab language (Python or shell) once.
-Provide starter files in `lab/` where useful and reference them by path.*
+*The only place the reader runs things. End at a keyboard, not a quiz. State the lab language
+(Python or shell) once; provide starter files in `lab/` and reference them by path. **From Module 4
+on, the learner directs the AI agent (Claude Code as the worked example) to do the git/setup work
+and then verifies it — they don't type the commands by hand.** In Modules 1–3 the learner still
+runs git manually, on purpose.*
 
 **You'll need:** *<tools/setup required for this lab>*