Catalog of UniverLab skills for AI agents.
The catalog is organized in three families, designed for dynamic, on-demand loading: skills are narrow and single-purpose so an agent (or a Canopy node pin) pulls only what the task needs. A Rust-loop skill in a Python session is context rot — install less, pin per role.
How the agent operates. execution-mindset is always on; the other two stack
on top per role.
Default operating mode: zero-indulgence clarity, interview when uncertain, the systemic resolution model (model → locate divergence → smallest decisive move → close the loop), verification before reporting, and token efficiency.
Behavior for design work: boundaries as contracts with violation signals, design for the process dying mid-step, reversible-vs-expensive decisions, and specs written for the cheapest executor.
Behavior for code work: structural judgment, pattern selection with restraint, maintainable multi-file changes, and debugging as the resolution model applied to code.
git clone https://github.com/UniverLab/skills.git
# or as a submodule
git submodule add https://github.com/UniverLab/skills.git skillsOnly installed where the canopy binary exists — these are inert without the
Canopy MCP server.
The Project Intelligence Layer: pull workspace context at session start,
persist facts/patterns/summaries with intelligence_upsert.
The action-driven workspace protocol: classify actions by risk and follow the declare → execute → report sequence.
Orientation map for "what can I do here?": capabilities, starter automation ideas, and how to discover the live tool surface without trusting a static list.
Design multi-step processes as reusable Canopy loops: the ROLE/WHAT/HOW spec contract, design-expensive/execute-cheap model selection, graph patterns, and the MCP tool playbook.
Bound to their CLI; installed automatically during Canopy setup only if the binary is available.
CLI to compile LaTeX documents to PDF without TeX Live, MiKTeX, or external
dependencies. Uses tectonic as the internal engine. Requires texforge.
CLI for CAD as code: declarative geometry in .cf files (TOML), compiled to
DXF with PNG preview. Requires cadspec.
CLI for making terminal demos as reproducible code. Capture, record, export
to gif/mp4. Requires demo.
# Manual installation of a single skill
npx skills add https://github.com/UniverLab/skills --skill texforge
# Or copy directly
cp -r skills/texforge ~/.agents/skills/With Canopy's dynamic skills, pin per node role instead of installing globally:
| Node role | Pin |
|---|---|
| Implementer | execution-mindset + code-engineering |
| Spec author / loop designer | architect-mindset + canopy-loop-design |
| Reviewer | execution-mindset (zero indulgence is the whole job) |
| Any Canopy-workspace session | canopy-intelligence, canopy-sync |
Each skill follows this structure:
skill-name/
├── SKILL.md # Main definition
├── README.md # Full documentation
├── LICENSE # MIT License
└── references/ # Reference documentation, loaded on demand
├── *.md # Practical guides and examples
Reference these skills in your agent configuration:
skills:
- name: execution-mindset
path: skills/execution-mindset/SKILL.md
always_active: true
- name: code-engineering
path: skills/code-engineering/SKILL.md
triggers:
- "write code"
- "review code"
- "design pattern"
- name: architect-mindset
path: skills/architect-mindset/SKILL.md
triggers:
- "design"
- "architecture"
- "write specs"
- name: canopy-loop-design
path: skills/canopy-loop-design/SKILL.md
triggers:
- "create loop"
- "plan loop"
- "orchestrate agents"Contributions welcome! Follow these guidelines:
- One skill per directory
- Clear documentation — Explain when and how to use the skill
- Practical examples — Include code and real scenarios
- MIT License — All contributions must be MIT
- Progressive disclosure — Keep
SKILL.mdfocused, move optional detail toreferences/ - Narrow scope — a skill that only applies next to a binary or MCP server declares it with
requires