stan-skills

Transparent, reproducible, deliverable Stan workflows — as Agent Skills. 透明、可复现、可交付的科研工作流。

A Claude Code plugin marketplace by wmj.

Plugin	What it does
stan-highlight	Colours Stan code by variable role, the block where each variable is declared, and outputs HTML for slides, Quarto notes and handouts.

---

Install

In Claude Code:

/plugin marketplace add perlatex/stan-skills
/plugin install stan-highlight@stan-skills

Then restart Claude Code (or run /reload-plugins). That's it, no dependencies beyond Python 3.8+.

Use it. The standard request is one sentence:

请使用 stan-highlight 高亮 model.stan，输出 model-highlighted.html

Use stan-highlight to highlight model.stan and write model-highlighted.html

If you leave out the output name, you still get model-highlighted.html next to model.stan. Other things you can ask:

高亮 model.stan，加上颜色图例

只输出 <pre> 片段，我要放进 Quarto 幻灯片

把 models/ 下所有 Stan 模型做成高亮 HTML，打包成 zip

You can also invoke it explicitly with /stan-highlight:stan-highlight model.stan.

Update later with /plugin marketplace update stan-skills.

Install as a plain skill (no plugin system)

Copy the skill folder into your personal skills directory:

# macOS / Linux
cp -r plugins/stan-highlight/skills/stan-highlight ~/.claude/skills/

# Windows (PowerShell)
Copy-Item -Recurse plugins\stan-highlight\skills\stan-highlight $HOME\.claude\skills\

The same folder works in other agents that read SKILL.md skills.

---

stan-highlight

Colour rules

Declared in	Colour
data	blue
transformed data	#0097A7
parameters	red
transformed parameters	orange
model	black bold
generated quantities	green

A variable keeps its colour everywhere it is used: N is blue inside model {} too, so students can trace where every quantity comes from.

These stay plain: comments, strings, types, function names, loop indices, and the whole functions {} block. Argument names there often reuse names declared elsewhere, so colouring them would mislead. Locals declared inside model or generated quantities are coloured only within their own block.

Why a script, not the model's judgement

The skill instructs Claude to run a deterministic Python parser (scripts/highlight_stan.py, standard library only) instead of colouring tokens by hand. Every run can be verified with --check:

1. Round trip: stripping the HTML tags gives back the exact source.
2. Completeness: every occurrence of a declared variable is coloured with its block's colour.
3. Loop indices: no loop index is ever coloured.

Command line (also works without Claude)

S=plugins/stan-highlight/skills/stan-highlight/scripts/highlight_stan.py

python3 $S model.stan                         # -> model-highlighted.html
python3 $S model.stan --legend                # with colour legend
python3 $S model.stan --fragment -o pre.html  # <pre> only, for Quarto / reveal.js
python3 $S models/ -r --zip --check           # whole tree -> models-highlighted/ + .zip

Batch mode mirrors the folder structure, so many files called model.stan never overwrite each other.

In Quarto

Styles are inline, so a fragment can be dropped into any HTML/reveal.js output:

```{=html}
<!-- contents of pre.html -->
```

or htmltools::includeHTML("pre.html") from an R chunk. PDF/LaTeX output is not covered yet.

Examples

See examples/: a hierarchical linear model and a latent class model (with a functions block and marginalised discrete class), each with the .stan source and the highlighted .html.

---

Repository layout

stan-skills/
├── .claude-plugin/marketplace.json        # this repo is a marketplace
├── plugins/
│   └── stan-highlight/
│       ├── .claude-plugin/plugin.json
│       └── skills/stan-highlight/
│           ├── SKILL.md
│           └── scripts/highlight_stan.py
├── examples/
└── docs/screenshot.png

License

MIT © wmj