Prior art#
Conda-skills was inspired by Pavel Zwerschke’s pixi-skills and skill-forge, which distribute agent instructions and supporting files through existing package managers. His post, Managing Agent Skills with Your Package Manager, describes the approach.
The Agent Skills format#
The underlying Agent Skills specification was originally developed by Anthropic and released as an open standard. It defines the skill directory, required metadata, instructions, and supporting resources. Its development is hosted in agentskills/agentskills.
The format and the distribution method are separate: Agent Skills describes the content, while Pavel’s work applies package managers to distributing it. Conda-skills adopts that distribution approach using conda’s Python APIs. See Skill package format for the distinction between specification requirements, package layout, and the checks performed here.
Bringing the approach to conda#
Conda-skills is an independent implementation built on conda’s plugin API, environment lookup, and installed package records. It does not bundle pixi-skills source code.
The projects share an installed package layout, but their discovery and validation rules differ. See How agents find skills for compatibility details and Skill package format for the requirements conda-skills checks.
The Proposal: package Agent Skills with conda draft proposes documenting a shared packaging convention. It credits this work as its starting point and leaves unresolved differences for discussion with the maintainers. It is not an official conda specification.
Complementary projects#
pixi-skills makes the package-manager approach available to Pixi users.
skill-forge provides examples and recipes for packaging skills.
conda-skills applies the approach to conda environments and the conda Python plugin ecosystem.
See Create a skill to start writing a skill, and Skill safety for the separate question of whether to trust its instructions.