---
title: "Using the MethScope agent skill"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Using the MethScope agent skill}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r, include = FALSE}
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)
```

## Overview

The MethScope repository includes an optional agent skill under
`agent-skills/methscope/`. The skill gives coding agents project-specific
context for MethScope package usage, documentation, `.cg` and `.cm` input files,
MRMP embeddings, cell-type prediction, model training, deconvolution,
visualization, input conversion, and the `methscope-cli` command-line
implementation.

Use the skill when you want an agent to follow the MethScope documentation and
repository conventions instead of giving generic methylation-analysis advice.

## Install for Codex

From the root of a cloned `zhou-lab/MethScope` repository, copy the Codex skill
entry point and shared MethScope context into your Codex skills directory.

```{bash, eval=FALSE}
mkdir -p "$CODEX_HOME/skills/methscope"
cp agent-skills/methscope/codex/SKILL.md "$CODEX_HOME/skills/methscope/SKILL.md"
cp -R agent-skills/methscope/core "$CODEX_HOME/skills/methscope/core"
```

The installed layout should look like this:

```text
$CODEX_HOME/skills/methscope/
  SKILL.md
  core/
    INSTRUCTIONS.md
    REFERENCES.md
    WORKFLOWS.md
```

After installation, start a new Codex session or reload the available skills.
Then ask Codex to use the MethScope skill for MethScope-specific tasks.

## Example Codex prompts

```text
Use the MethScope skill to help me test GenerateInput with inst/extdata/example.cg.
```

```text
Use the MethScope skill to explain how to convert an ALLC file to a MethScope .cg file.
```

```text
Use the MethScope skill to debug a PredictCellType shape mismatch.
```

```text
Use the MethScope skill to update the deconvolution tutorial for bundled reference matrices.
```

## Install for Claude-style project context

For Claude or other agents that use repository-level project instructions, use:

```text
agent-skills/methscope/claude/CLAUDE.md
```

If your project already has a `CLAUDE.md`, copy the MethScope-specific guidance
from that file into your existing instructions. If your project does not already
have one, use `agent-skills/methscope/claude/CLAUDE.md` as the starting project
context.

Keep the shared files in `agent-skills/methscope/core/`, because the Claude
instructions point to the same package-specific references used by the Codex
skill.

## What the skill knows

The shared MethScope skill context includes:

- The difference between GitHub full example data and CRAN toy data
- The role of `.cg`, `.cg.idx`, `.cm`, and bundled reference `.rds` files
- Standard workflows for `GenerateInput()`, `PredictCellType()`,
  `Input_training()`, `PlotUMAP()`, `PlotConfusion()`, `PlotF1()`, and
  `nnls_deconv()`
- Guidance for using `example_label.csv` with `.cg.idx` sample order
- Input conversion pointers for BED-like, ALLC, beta/fraction, and binary tracks
- The relationship between the MethScope R package and `methscope-cli`
- `methscope-cli` bundle types: `.ubjx`, `.refx`, and `.updecx`

## Update the skill

When MethScope documentation or package behavior changes, update the source
files in:

```text
agent-skills/methscope/
```

Then reinstall or recopy the skill into your local agent configuration. The
repository copy is the source of truth.