Agents

Agents Skills

Skills are reusable instruction packs that keep your agents consistent across tasks. Pair this with the agents overview to design composable, maintainable workflows.

Skills complement MCP tools by wrapping instructions and context around raw API calls. Use skills for higher-level orchestration and MCP for direct tool execution.

Structure

Use a predictable folder shape so skills are easy to audit and evolve.

Directory layout

SKILL.mdmarkdown

Main instruction file. Keep it short and task-focused so agent selection stays fast.

references/directory

Supporting docs like request contracts, troubleshooting guides, and API references. Linked from SKILL.md, not embedded.

scripts/directory

Deterministic scripts for repetitive transformations, validation, and data normalization.

Execution Model

Skills should reduce ambiguity, not add another reasoning burden.

Keep SKILL.md short

Keep the main SKILL.md short and task-focused so selection stays fast. Agents scan skill files to pick the right one.

Link, don't embed

Link supporting references instead of embedding long implementation dumps. This keeps the primary instruction lean.

Prefer deterministic scripts

Use scripts for repetitive transformations and validation. Deterministic code is more reliable than repeated LLM inference.

Define success criteria

Define explicit success criteria to keep multi-agent handoffs consistent. Each skill should know when it's done.

Versioning

Treat skills like interfaces: stable, explicit, and testable.

Tag behavior changes

Track breaking prompt or output contract changes in a version note section. Consuming agents may depend on specific output formats.

Preserve backwards compatibility

When possible, support both old and new parameter names during a migration window. This avoids breaking existing agent workflows.

Test representative tasks

Run a stable set of before/after agent tasks whenever a skill is updated. Catch regressions early.

Install

Install from the official public repository by explicit skill name.

Install command

npx skills add https://github.com/Theoreumont/nativeform-skills --skill nativeform-agent-workbench