Skills
Skills are atomic units of domain expertise that agents load on demand. They teach the agent how to use a Python package, query a database API, or apply a scientific methodology -- without being tied to a specific workflow. A skill is stored as a Markdown file, usually with optional bundled scripts, examples, or package references alongside it.
How skills work
Rather than packing every possible piece of domain knowledge into the agent's system prompt, Beakr uses a just-in-time loading model:
- Catalog. The agent sees a compact catalog of available skills -- name, description, and trigger keywords -- in its system prompt.
- Match. When the user's request matches a skill's trigger keywords or domain, the agent identifies which skill to load. Users can also ask for a specific skill by name when they want to force a known workflow.
- Load. The agent calls
load_skillsto pull the full expertise into context -- package documentation, code patterns, API conventions, and bundled scripts. - Execute. The agent applies the loaded knowledge to the user's task, using the correct packages, APIs, and methodologies.
This keeps the agent's context focused. A skill for single-cell RNA-seq analysis doesn't consume context when the user is asking about project management.
What a skill contains
| Component | Purpose |
|---|---|
| Name and description | Shown in the skill catalog. The agent uses these to decide when to load the skill. |
| Markdown instructions | The core skill content. Markdown keeps procedures portable, reviewable, and easy to share across Beakr-supported agent surfaces. |
| Domain | Knowledge domain (e.g., bioinformatics, cheminformatics, data science). Used for categorization. |
| Trigger keywords | Comma-separated keywords that trigger automatic matching (e.g., "scRNA-seq, h5ad, UMAP, Scanpy"). |
| Knowledge | The full expertise content: how to use the package, API conventions, best practices, common pitfalls. |
| Packages | Python packages to install in the sandbox when this skill is active (e.g., ["scanpy", "anndata"]). |
| Tool patterns | Guidance on how to use Beakr's built-in tools for this domain. |
| Bundled scripts | Pre-built code scripts the agent can reference or execute directly. |
Skill ownership
Skills can come from three sources:
| Source | Owned by | Visible to |
|---|---|---|
| System | Beakr | All users |
| Organization | Your org | All members of the org |
| User | Individual user | Only that user |
Organizations can create custom skills that encode their specific methodologies, internal tools, or proprietary workflows. These skills are available to all members of the organization but not visible to other organizations. Because the source format is portable Markdown, approved skills can be shared across teams and reused anywhere Beakr agents run.
Skill library
Each user has a personal skill library -- a set of installed skills that are automatically available in their agent sessions. Users can:
- Browse the skill catalog and install skills they need
- Pin skills to specific conversations or workflows
- Dismiss skills they don't use
- Request new skills from their organization admin
Skills vs. workflows
A skill teaches the agent how to do something (use Scanpy, query PubMed, apply a statistical method). A workflow tells the agent what to do (run this analysis pipeline on this data). Workflows reference skills -- the workflow provides the steps, the skill provides the expertise for each step.