Skills - LatchBio

Skills are prompt packages that teach the agent new capabilities — domain-specific workflows, analysis templates, institutional conventions, or technology-specific guidance. Each skill is a directory with a SKILL.md entrypoint that the agent loads on demand based on context.

Developing Skills (Branch Workflow)

You can test skill changes on a branch before merging to main by pinning a notebook to a specific latch-skills branch.

Create a branch

Push a branch to the latch-skills repository with your changes (new or modified skills).

Set the skills branch

In Edit mode, open Runtime Settings (via the runtime dropdown in the notebook top bar) and select your branch from the Skills Branch picker.

Restart the pod

Restart the notebook runtime. On startup, the agent clones latch-skills at your chosen branch instead of main.

Test

Send a prompt that should trigger your skill. Verify the agent loads and uses it correctly.

Merge and reset

Once verified, merge your branch into main. Clear the Skills Branch setting (or leave it as default) so the notebook tracks main going forward.

The skills branch setting is per-notebook and persists across restarts. On every restart, the runtime fetches the latest commit on the configured branch, so you can push updates and restart without changing the setting.

Built-in Skills

Every agent session includes Latch’s public skill set from latch-skills. These are cloned automatically at startup — no registration required. By default, the main branch is used. See Developing Skills above to test on a different branch.

Platform Skills

Core Latch Plots APIs and patterns:

Skill	Description
`latch-plots-ui`	Widget APIs for output (plots, tables, H5 viewers), user input (selects, sliders, checkboxes), and layout (rows, columns, grids)
`latch-data-access`	File selection widgets, Latch Data browsing, Registry tables, and LPath utilities for `latch://` paths
`latch-workflows`	Launching and monitoring bioinformatics workflows — parameter construction, validation, and output retrieval
`latch-curation`	Curating external datasets (GEO/GSE) into Latch-compatible AnnData with Ensembl gene IDs and ontology-annotated metadata

Technology Skills

Provider-specific analysis workflows and best practices:

Skill	Technology	Description
`takara-devkit`	Takara Seeker / Trekker	QC, background removal, normalization, clustering, differential expression, and cell typing
`xenium-devkit`	10x Xenium	Data preparation, preprocessing, differential expression, cell type annotation, and domain detection
`vizgen-devkit`	Vizgen MERFISH	Cell segmentation, preprocessing, QC, spatial analysis, and secondary analysis
`atlasx-devkit`	AtlasXomics DBiT-seq	QC, clustering, differential analysis, and cell type annotation for spatial ATAC-seq

Custom Skills

Organization admins can register private GitHub repositories containing custom skills that get loaded into agent runtimes across all workspaces in the organization.

Skill Repository Structure

A skill repository must follow this structure:

my-org-skills/
├── README.md                     # For humans browsing GitHub (not loaded by agent)
│
├── spatial-analysis/             # Each top-level directory = one skill
│   ├── SKILL.md                  # Required: skill entrypoint
│   ├── reference.md              # Optional: supporting documentation
│   └── examples/
│       └── xenium-workflow.md    # Optional: example outputs
│
├── curation/
│   ├── SKILL.md
│   ├── steps/
│   │   ├── harmonize.md
│   │   └── download.md
│   └── scripts/
│       └── validate.py           # Optional: scripts the agent can execute
│
└── volcano-plot/
    ├── SKILL.md
    └── templates/
        └── template.md           # Optional: templates for the agent to fill in

Rules

Each top-level directory containing a SKILL.md file becomes a skill
Only one level of skill directories — no nesting skills inside skills
README.md at the repo root is for GitHub — the agent does not load it
Skill directory names must be unique across all registered repos in an organization. If two repos define the same skill name, syncing will fail.

Writing a SKILL.md

Every skill needs a SKILL.md file with YAML frontmatter and markdown instructions:

---
name: spatial-analysis
description: >
  Spatial transcriptomics analysis workflows. Use when the user is
  working with Xenium, Visium, or MERFISH data and needs help with
  spatial analysis, neighborhood detection, or co-expression.
---

When performing spatial analysis:

1. Identify the technology platform from the data
2. Load the appropriate reference panel
3. Follow the platform-specific QC workflow
4. Generate spatial plots with appropriate coloring

For detailed API usage, see [reference.md](reference.md).

Required Fields

Field	Description
`name`	Lowercase letters, numbers, and hyphens only (max 64 characters). Used as the skill identifier.
`description`	What the skill does and when to use it. The agent reads this to decide when to load the skill automatically.

Optional Fields

Field	Description
`allowed-tools`	Restrict which tools the agent can use (e.g., `Read, Grep, Glob`).
`disable-model-invocation`	Set to `true` to prevent auto-loading. The skill can only be triggered explicitly.
`argument-hint`	Hint shown for expected arguments (e.g., `[dataset-name]`).

Supporting Files

Keep SKILL.md focused and under 500 lines. Move detailed reference material to separate files and link them from SKILL.md:

## Additional resources

- For complete API details, see [reference.md](reference.md)
- For usage examples, see [examples/](examples/)

The agent loads supporting files on demand when it follows a link from SKILL.md.

Merging Multiple Repositories

When multiple repositories are registered with the same target path (the default .claude/skills/), their skill directories are merged side by side:

.claude/skills/
├── xenium-qc/          # from repo A
├── lab-conventions/     # from repo A
├── my-custom-plots/    # from repo B
└── curation/           # from repo B

Skill directory names must be unique across all repos. If two repos both define a spatial-analysis/ skill, syncing will fail and the conflict will be reported in Organization Settings → Agent Skills.

Registering a Repository

Create a GitHub Personal Access Token

Generate a GitHub PAT with repo scope so the agent runtime can clone private repositories.

Connect your PAT

Go to Organization Settings → Agent Skills → GitHub Authentication and enter your GitHub username and token.

Add a repository

Click Add Repository and provide:

Display Name: A human-readable label (e.g., “Our Lab Skills”)
Repository URL: The HTTPS clone URL (e.g., https://github.com/my-org/agent-skills.git)
Branch: The branch to track (defaults to main)
Target Path: Where to mount in the runtime (defaults to .claude/skills)

Verify

When a new workspace agent session starts, the registered repositories are cloned into the runtime. The agent discovers all skills automatically.

How Skills Are Loaded

When an agent runtime starts in a workspace belonging to your organization:

All registered repositories are cloned using the org’s GitHub PAT
Skill directory names are checked for uniqueness — conflicts block the sync
The agent discovers SKILL.md files and reads their description fields
During a session, the agent automatically loads a skill when your request matches its description

Skills do not need to be invoked explicitly — the agent decides when they are relevant based on the conversation.

Example: Technology-Specific Skill

---
name: xenium-qc
description: >
  Xenium spatial transcriptomics QC workflow. Use when the user has
  Xenium data and needs quality control, transcript filtering, or
  cell segmentation validation.
---

## Xenium QC Workflow

1. Load the Xenium output bundle with `xeniumranger`
2. Check transcript counts per cell (expect >50 median)
3. Validate cell segmentation boundaries
4. Filter low-quality cells (< 20 transcripts)
5. Generate spatial QC plots

For platform-specific thresholds, see [reference.md](reference.md).

Example: Institutional Conventions Skill

---
name: lab-conventions
description: >
  Lab-specific analysis conventions and standards. Use when generating
  plots, writing methods sections, or exporting results.
---

## Plot Standards

- Use colorblind-safe palettes (viridis, cividis)
- Font size: 12pt for labels, 10pt for ticks
- Always include scale bars on spatial plots
- Export as both PNG (300 DPI) and SVG

## Methods Sections

When writing methods text, cite the following versions:
- scanpy 1.10.x
- squidpy 1.4.x
- anndata 0.10.x

​Developing Skills (Branch Workflow)

​Built-in Skills

​Platform Skills

​Technology Skills

​Custom Skills

​Skill Repository Structure

​Rules

​Writing a SKILL.md

​Required Fields

​Optional Fields

​Supporting Files

​Merging Multiple Repositories

​Registering a Repository

​How Skills Are Loaded

​Example: Technology-Specific Skill

​Example: Institutional Conventions Skill

Developing Skills (Branch Workflow)

Built-in Skills

Platform Skills

Technology Skills

Custom Skills

Skill Repository Structure

Rules

Writing a SKILL.md

Required Fields

Optional Fields

Supporting Files

Merging Multiple Repositories

Registering a Repository

How Skills Are Loaded

Example: Technology-Specific Skill

Example: Institutional Conventions Skill