Skills and Resources¶

Skills package reusable agent capabilities. A skill can contain instructions, invocation metadata, input and output contracts, references, assets, and executable scripts.

PromptScript resolves each skill once and emits the target-native representation supported by each configured AI platform.

Inline Skill¶

@skills {
  security-review: {
    description: "Review code for application security risks"
    userInvocable: true
    disableModelInvocation: false
    allowedTools: ["Read", "Grep", "Bash"]
    content: """
      Review authentication, authorization, secret handling, input validation,
      dependency risk, and unsafe command execution.
    """
  }
}

Directory Skill¶

Portable skills can use a complete directory:

.promptscript/skills/security-review/
├── SKILL.md
├── LICENSE
├── references/
│   └── threat-model.md
├── scripts/
│   └── scan.sh
└── assets/
    └── report-template.md

SKILL.md frontmatter declares resources and scripts:

---
name: security-review
description: Review code for application security risks
license: MIT
compatibility: Requires shell access
metadata:
  owner: security
allowed-tools: 'Read Grep Bash'
references:
  - references/threat-model.md
scripts:
  - scripts/scan.sh
inputs:
  path:
    type: string
    description: Path to review
outputs:
  report:
    type: string
    description: Review report
---
Review selected code and use bundled references when evaluating risk.

PromptScript preserves the directory structure when producing native skill files. Script invocation and executable file modes remain platform and operating-system specific.

Skill Capabilities¶

model sets the model used while the skill is active. Claude Code and Grok Build write it to the skill file, mapped through the model catalog the same way as agent models. Other targets have no skill model field and leave it out.

For skills backed by a SKILL.md file, a model set in .prs wins over the SKILL.md frontmatter, and a frontmatter model without a .prs one is mapped the same way. A model the target cannot run is left out of the skill file and reported with a PS4004 warning during compilation.

Universal and Native Paths¶

PromptScript can discover shared skills from the universal .agents/skills/ directory:

universalDir: .agents

Target formatters then emit native paths such as:

• .claude/skills/<name>/SKILL.md
• .github/skills/<name>/SKILL.md
• .factory/skills/<name>/SKILL.md
• .opencode/skills/<name>/SKILL.md
• .agents/skills/<name>/SKILL.md

Gemini full mode emits interoperable .agents/skills/<name>/skill.md files alongside native commands under .gemini/commands/.

Filter Skills Per Target¶

id: filtered-skills-project
syntax: '1.4.0'

targets:
  - factory:
      version: full
      includeSkills:
        - security-review
        - release
  - claude:
      version: full
      includeSkills: true

Use skillBaseDir when generating skills inside a plugin, package, or nested monorepo build:

builds:
  payments:
    output: packages/payments
    targets:
      - factory:
          version: full
          skillBaseDir: .factory/skills

Import Skills¶

Skills can come from local files, registries, or Git repositories:

@use ./skills/security-review.md
@use @company/skills/release@^2.0.0
@use github.com/acme/agent-skills/[email protected]

Remote references are recorded in promptscript.lock with commit and integrity information.

Related Documentation¶

• Building Skills
• Local Skills
• Skill Composition
• Skill Contracts
• Shared Resources
• Language Reference: @skills