Bring Your Own Skill (BYOS)
When to Use
- Creating a new team/project skill that will be shared via git
- Assigning a project skill to a specific PRISM agent
- Scaffolding a new skill with the correct directory structure
- Validating existing project skills for correctness
How It Works
Claude Code Discovery (native)
Project skills live at .claude/skills/{skill-name}/SKILL.md. Claude Code discovers them automatically - no registration, sync, or hooks required. They take precedence over user-level and plugin skills.
PRISM Skill Discovery (existing infrastructure)
Skills opt into PRISM discovery via prism: frontmatter metadata. All skills with a prism: block are injected into every workflow step for every agent. The agent field is optional — it serves as informational metadata about which agent the skill was designed for, not as a filter.
---
name: my-team-skill
description: What this skill does
prism:
agent: dev # optional: sm | dev | qa | architect (informational hint)
priority: 10 # lower = higher priority (default: 99)
---
At runtime, discover_prism_skills() scans .claude/skills/*/SKILL.md and injects all discovered skills into every workflow step in priority order. Agents are instructed to ALWAYS prefer using an available skill over solving without one.
Quick Start
Scaffold a new skill
/byos scaffold my-skill --agent dev
Creates .claude/skills/my-skill/ with a pre-filled SKILL.md and /reference/ directory.
Validate a skill
/byos validate my-skill
Checks structure, frontmatter, prism: metadata, and token budget.
List project skills
/byos list
Shows all project-level skills with their agent assignments.
Reference Documentation
- Getting Started - Step-by-step guide for creating your first project skill
- Skill Template - Copy-paste ready SKILL.md template with all fields
- Examples - Real-world project skill examples with agent assignment
Guardrails
- Follow the 3-level pattern: metadata (~100 tokens), body (<2k tokens), reference files (unlimited)
- All reference
.mdfiles MUST go in/reference/- never in the skill root - Valid agents:
sm,dev,qa,architect - Skill names must be kebab-case (lowercase, hyphens only)
- One SKILL.md per skill - the only
.mdfile allowed in the skill root - For deep-dive skill authoring guidance (progressive disclosure, token optimization), use
/skill-builder