Glossary¶

Definitions follow the ISO 11179 metadata registry standard: precise, concise, distinct, non-circular, and free of business rules.

Agent Skill: A directory containing a SKILL.md file and optional supporting resources that teaches an AI agent how to perform a specific task through structured instructions.
Activation: The process by which an agent loads a skill's full SKILL.md body into context after determining relevance from the skill's description.
allowed-tools: An optional YAML frontmatter field specifying which tools a skill is permitted to use during execution. Experimental.
Assets Directory: An optional assets/ subdirectory within a skill containing static files used in output — templates, images, schemas, fonts.
Bloom's Taxonomy: A six-level classification of cognitive objectives (Remember, Understand, Apply, Analyze, Evaluate, Create) used to design learning outcomes and quality scoring rubrics.
Checkpoint: A designated point in a multi-skill pipeline where execution pauses for user review before proceeding to the next stage.
Context Window: The maximum number of tokens an AI model can process in a single session, shared across system prompt, conversation history, skill instructions, and generated output.
DAG (Directed Acyclic Graph): A graph structure where edges have direction and no cycles exist. Used for learning concept dependencies and skill pipeline ordering.
Description Field: The required YAML frontmatter field (max 1,024 characters) that tells agents what a skill does and when to use it. Loaded into the skill registry at session start.
Discovery: The startup phase where an agent scans skill directories and loads only the name and description of each available skill into its registry.
Frontmatter: The YAML metadata block at the top of a SKILL.md file, delimited by ---, containing required fields (name, description) and optional fields (license, compatibility, metadata, allowed-tools).
ISO 11179: An international standard for metadata registries that defines five qualities of good definitions: precise, concise, distinct, non-circular, and free of business rules.
Lazy Loading: A token efficiency pattern where reference documents are read into context only when a specific workflow step requires them, rather than at skill activation.
MCP (Model Context Protocol): An open protocol that tells agents what tools exist and how to call them. Complementary to skills: MCP provides tool discovery, skills provide workflow knowledge.
Meta-Skill: A skill that routes requests to one of several sub-skills based on keyword matching. Uses a references/ directory containing guide files for each variant.
MicroSim: An interactive browser-based simulation built with JavaScript libraries (p5.js, Chart.js, vis-network) for educational visualization.
Name Field: The required YAML frontmatter field (max 64 characters) that identifies a skill. Must be lowercase with hyphens only and match the parent directory name.
Pipeline: A sequence of skills executed in dependency order where each skill's output serves as the next skill's input.
Progressive Disclosure: The three-tier architecture for managing skill token costs: Tier 1 metadata (~100 tokens) loaded always, Tier 2 instructions (<5,000 tokens) loaded on activation, Tier 3 resources loaded on demand.
Quality Gate: A threshold-based checkpoint within a skill where output is scored against a rubric and execution continues only if the score meets the minimum (typically 70-85 out of 100).
Quality Scoring Rubric: A structured evaluation framework within a skill that decomposes output quality into weighted criteria summing to 100 points, with defined thresholds for proceed/stop decisions.
References Directory: An optional references/ subdirectory within a skill containing documentation that agents load on demand during execution.
Routing Table: A mapping within a meta-skill that associates keywords or request patterns with specific reference guide files in the references/ directory.
Scripts Directory: An optional scripts/ subdirectory within a skill containing executable code (Python, Bash, JavaScript) that agents run during workflow steps.
Session Logging: The practice of writing structured log files to a logs/ directory recording timing, files created, quality scores, and decisions made — enabling cross-session continuity.
Skill Collection: A repository containing multiple related skills packaged together with a shared install script and documentation.
Skill Registry: The in-memory index an agent builds at startup from all available skill names and descriptions, used to match incoming requests to relevant skills.
SKILL.md: The required markdown file defining a skill's metadata, instructions, and workflow. Contains YAML frontmatter followed by structured markdown content.
Skip-if-Complete: A token efficiency pattern where a workflow step checks for existing output before executing, avoiding redundant work when resuming from a prior session.
Step 0: The conventional first step in a skill workflow dedicated to environment setup — detecting project context, validating prerequisites, and prompting for missing information.
System Prompt Budget: The portion of the context window allocated to system-level instructions, including all loaded skill descriptions. The 30-skill limit exists to keep this budget manageable.
30-Skill Limit: The maximum number of skills that can be loaded into a single Claude Code session, determined by the system prompt token budget.
Token Efficiency: The practice of minimizing unnecessary token consumption within skills through progressive disclosure, lazy loading, skip-if-complete detection, and concise instruction writing.
User Dialog Trigger: A condition defined within a skill workflow that causes the agent to pause and ask the user a question before proceeding — typically at quality gates, destructive actions, or ambiguous decision points.
Workflow Step: A numbered section within a SKILL.md body containing specific instructions for one logical unit of work within the skill's execution sequence.