Glossary

Defines the core terms used by Skillset's source, build, provider, and activation workflows.

Adaptive source

Repository-owned Skillset source that expresses a shared authoring intent and can be transformed into one or more faithful provider-native destinations. Adaptive does not mean every provider supports every feature.

Activation

A provider or user action that makes generated material discoverable, trusted, enabled, or connected at runtime. Activation is outside ordinary Skillset build operations.

Agent standard

A provider-neutral format that Skillset renders inherently when the workspace has applicable source: Agent Instructions (AGENTS.md), Agent Skills (.agents/skills/), and Agent Plugins (plugins/<plugin>/). Standards are not targets and have no selector in configuration; compile.targets chooses providers only. Each standard is adopted through a standard profile.

Baseline

The portable output an Agent standard produces from source, such as a .agents/skills/<skill>/SKILL.md. A provider that can read the same file adds a delta instead of writing its own copy. skillset diff and skillset explain label this as, for example, agent-skills baseline.

Build

The deterministic derivation and rendering of validated Skillset source into repo-local provider-native files. Builds preview by default and write only with explicit authority.

Canonical source

The one authored location that owns a fact or source unit. Generated projections and explanatory summaries refer back to that owner instead of becoming competing truth.

Cascade

The scoped inheritance of workspace and plugin defaults into a source unit, with narrower explicit configuration taking precedence. Cascade does not mean arbitrary filesystem inheritance.

Consumer

A logical projection that a generated file represents: a standard baseline or a provider delta. One physical file has one owner and can have several consumers; for example, .agents/skills/ serves both the Agent Skills baseline and the Codex delta. The lock records every consumer. A consumer is provenance, not proof that a provider has discovered or activated the file.

Delta

The provider-specific part a target contributes on top of a compatible baseline, instead of writing a separate file. skillset diff shows a shared file as, for example, [agent-skills baseline + codex delta].

Destination

One concrete output location and format owned by a target adapter, such as a Claude skill directory, a Codex AGENTS.md, or a Cursor rule file.

Drift

A difference between current canonical source and its expected generated output, or between a generated contract and its checked-in documentation projection.

Generated output

Provider-native files derived from Skillset source. Generated output may be committed and reviewed, but it is disposable and is not the authored source of truth.

Loadout

A reusable collection of agent-facing skills, rules, agents, hooks, plugins, and related resources prepared for one or more providers.

Lock

A generated skillset.lock file beside managed output. It records source paths, generated paths, hashes, owners, consumers, and projection roles, so checks can tell managed output from unrelated files. Locks are generated output; never edit them by hand.

Owner

The single standard or provider recorded as responsible for writing one physical file in a standard or provider projection. Bundle entries in a lock, such as preserved target-native islands, carry no owner or consumers. When a standard and a provider share a file, the standard owns it and the provider is a consumer. skillset explain prints this as, for example, owner: agent-skills baseline.

Project agent

A repository-scoped subagent authored under .skillset/subagents/ and rendered to each provider’s agent location, such as .claude/agents/. Project agents are source; provider settings for them, such as a default model, live under defaults.<provider>.agents.

Projection

A deterministic set of generated artifacts derived from canonical source or a typed contract. A projection is checked for freshness but is not edited as source truth.

Provider-native

Data or files shaped for a provider’s own supported interface rather than forced into a lowest-common-denominator format.

Render

To transform validated source intent into one target’s provider-native destination shape without installing or activating the result.

Render result

A structured record of what one renderer emitted, skipped, degraded, or could not support, including the evidence needed for diagnostics and provenance.

Rule

Authored agent guidance. The front page, .skillset/RULES.md, renders to Claude’s project instruction file and the first section of root AGENTS.md. Scoped rules under .skillset/rules/**/*.md can be limited to paths and render to Claude rules, Cursor rules, and root or scoped AGENTS.md files. AGENTS.md output is owned by the Agent Instructions standard. See the Rules reference.