repo·evals
· 2026-05-05 ·main@HEAD

codebase-to-course

zarazhangrui/codebase-to-course

🛠60 / 100
✅
⚠
🎯
⚠

⚛
→
⚗
→
🧬

🛑
0–29
⚠️
30–49
🛠
50–79
🏭
80–100
▼
60
🛠· 60 / 100
  • ✓5 claims passed, no critical failures
  • ⚠README may claim a license but no LICENSE file exists
  • ◐release_pipeline=1, recently_active=True
  • ⚪EN-only or ZH-only README
  • ⚪static-only eval; live e2e pending

#1👤
#2🎯
#3🧭
#4⇄

Your codebase你的代码库(any language)(任意语言)LLM reads codeLLM 读代码→ module briefs→ 模块大纲Per-module HTML每模块 HTML+ quizzes + diagrams+ quiz + 图示build.shbuild.shstitches index.html拼成 index.htmlSingle-page单页交互课程interactive course

Copy codebase-to-course/ to ~/.claude/skills/anyeasy
  • 📡
Anthropic Claude (Claude Code)
LLM analyzes codebase + writes module briefs
Standard Claude Code cost; per-course token cost depends on codebase size
Google Fonts CDN
Course typography
Optional; course works offline if fonts are cached
· 7
5 1 1
+40
+10
+5
+3
+5
-3

6 / 7
passed claim-001

passed claim-002

passed claim-003

passed claim-004

passed claim-005

failed claim-006

untested claim-007

input_contract
output_contract
determinism
idempotence
no_skill_callouts
failure_mode_clarity

  • evidence_completeness='partial' (not portable) → capped at 'usable'

  • only 2/3 critical claims covered

archetype: hybrid-skill→core_layer_tested? True→evidence: partial→recommended: usable→final: usable
ceiling 1 · evidence_completeness='partial' (not portable) → capped at 'usable'
claim-001SKILL.md frontmatter 完整且有 trigger 短语清单criticalskill-shape● passed
claim-002references/ 真实 ship 了 CSS / JS / build.sh，不只是 markdowncriticalskill-completeness● passed
claim-003设计系统 / 内容哲学 / interactive-elements 三份文档真实有内容highdepth● passed
claim-004build.sh 是真实可执行的 assemblerhighassembly-tooling● passed
claim-005目录结构与 README 描述一致highcontract-alignment● passed
claim-006仓库有 LICENSE 文件，下游可二次使用highlicensing✕ failed
claim-007端到端 happy path：对一个真实代码库跑出一个能开的课程criticalend-to-end○ untested

0%
0.00s
0

run-static-checks

2026-05-05
0% — tokens in ? / out ?

run-static-checks

2026-05-05
0% — tokens in ? / out ? 
# codebase-to-course — final verdict (2026-05-05)

## Repo

- **Name:** zarazhangrui/codebase-to-course
- **Branch:** main@HEAD
- **Archetype:** hybrid-skill
- **Layer:** molecule
- **Stars:** 4,224

## What was evaluated

| Claim | Status | Notes |
|---|---|---|
| 001 SKILL.md frontmatter + triggers | passed | 7+ trigger phrases in description; auto-discovery should work |
| 002 real shipped assets | passed | styles.css 1195 lines + main.js 498 lines + build.sh + 2 HTML templates |
| 003 reference docs depth | passed | 5 docs all > 2 KB; interactive-elements.md = 32 KB (deepest) |
| 004 build.sh is real assembler | passed | `set -e` + cat _base + modules + _footer → index.html |
| 005 directory matches README | passed | structure 1:1 with documented layout |
| 006 LICENSE present | **failed** | HTTP 404 — no LICENSE file at root |

| 007 live agent end-to-end | untested | needs Claude Code session on a real codebase |

## Real findings

1. **No LICENSE file.** A 4.2K-star skill users copy into their
   personal `~/.claude/skills/` should have one. Without it, anyone
   modifying / forking / commercially distributing the generated
   courses has no legal cover. **Easy fix upstream:** add MIT or
   Apache-2.0.

2. **Genuinely hybrid, not "markdown with hopes".** Many skills label
   themselves "hybrid" but ship only SKILL.md. This one ships 1,200
   lines of CSS, 500 lines of JS, an assembler script, and 5
   substantive reference docs (totaling ~57 KB). The LLM has real
   building blocks to work from.

3. **`build.sh` is minimal but correct.** 6 lines. Could be done
   inline in SKILL.md as a code block, but having it as a script
   means users can re-run `bash build.sh` after editing modules.

4. **interactive-elements.md is the deepest spec (32 KB).** Sensible
   prioritization — quizzes / animations / visualizations are the
   most variable surface; design system + content philosophy are
   shorter because they need fewer worked examples.

## Why not higher

- Single live run not yet logged (molecule layer cap).
- LICENSE missing — small but real legal gap that pulls the score.

## Path forward

1. Add LICENSE (MIT or Apache-2.0) → claim-006 → passed → +5 score.
2. Run on a real codebase in Claude Code → claim-007 → passed → live
   evidence raises confidence to high.
3. Multiple logged runs across codebase shapes → score 80+.

## Recommended

```yaml
status: evaluated
```

The score model will assign the bucket; this dossier is the
evidence trail behind it.