agent-teacher: Instructions file for Claude Code

AgentTeacher

Concept-teaching skill: explains a technical concept through a fixed six-part structure — intuition → example (runnable or pseudocode) → walkthrough → trap → pointers → test questions.

Before you start

• Personal / global agent rules belong outside this repo. This file records only AgentTeacher's Claude Code entry and maintenance rules.
• The full repository map, working rules, current risk areas, verification, and release flow live in AGENTS.md.
• The skill itself starts at SKILL.md. Teaching method: references/teaching-method.md. Code style for teaching: references/code-style-for-teaching.md. Concept→language mapping: references/concept-to-language.md. Enrichments: references/enrichments.md.
• For contributors (English / 中文 onboarding): see CONTRIBUTING.md / CONTRIBUTING.zh-CN.md.

Common commands

bash scripts/package-skill.sh                  # build dist/agent-teacher.zip

The project has no build / render pipeline — output is markdown delivered in chat. There is no test runner; verification is done by running the skill on a real prompt and inspecting the lesson.

Project-specific hard rules

• The six-part spine (intuition → example → walkthrough → trap → pointers → test questions) is the contract. Don't add a seventh part, don't drop parts. If one feels redundant for a specific concept, write a one-liner — don't skip.
• Mode A (runnable) vs Mode B (pseudocode) is binary. Mode B auto-writes a sidecar file at /tmp/<slug>-pseudocode.<ext>; Mode A writes a file only when the user asks.
• Pseudocode must be real Python / PyTorch syntax with state annotations on every line where state changes. Natural-language pseudocode (FOR each token DO ...) is forbidden — it throws away the precision code form gives.
• The test questions section is not a comprehension check. Hints, no answers. Phrase each question concretely ("if all rewards in a group were identical, what happens to the gradient?" — not "do you understand?"). Don't frame as interview questions; don't use the word "interview" in skill output.
• The eight enrichments (math tier, state tracking, design rationale, cousin matrix, prerequisites + variants, misconception list, visualization, invariants) are concept-driven, not concept-default. Decision table is in references/enrichments.md. Hard cap 5 per lesson, hard floor 0 (plain six-part lesson is the right answer for most concepts). The framework was tuned for DL/ML but state tracking / design rationale / cousin matrix / prereqs+variants / visualization / invariants also apply to data structures, distributed protocols, and complex algorithms. Language features (closures, generators, channels) never get enrichments.
• Mode B has three flavors: DL (tensor shape annotations), protocols (participant state + message arrows), algorithms / data structures (invariants + concrete trace). State tracking means the right thing in each flavor — don't apply DL shape-annotation conventions to a protocol pseudocode block.
• When changing SKILL.md, sync the matching playbook in references/teaching-method.md. SKILL.md is the contract; references are the expanded playbook with examples. They must agree.
• Project canonical docs are English-only (SKILL.md, AGENTS.md, CLAUDE.md, references/). Bilingual surfaces are README and CONTRIBUTING (both in English and Chinese versions) and assets/examples/ (each example exists as -en.md and -zh.md). See CONTRIBUTING.md for the full bilingual policy.
• Don't bundle evals/ / .claude-plugin/ / dist/ into the release zip. The exclude list lives in scripts/package-skill.sh.
• After editing scripts/package-skill.sh or adding any file that should ship in the zip, refresh and inspect dist/agent-teacher.zip, confirming new files were git add-ed (the package script is based on git ls-files; untracked files vanish silently).
• Don't commit one-off review reports or diagnostic snapshots. Stable rules go into AGENTS.md, CLAUDE.md, SKILL.md, or references/; everything else is discarded.