Docs

Agent Skills

NanoCodana supports Anthropic-style Agent Skills: a SKILL.md with frontmatter and instructions that the model loads on demand. The agent always sees a short catalog of names and descriptions; a skill's full instructions enter context only when a task matches.

Where skills live#

Skills are auto-discovered from the agent's own filesystem at the conventional directories — identically on every platform:

.agents/skills/<name>/SKILL.md     # cross-agent convention
.claude/skills/<name>/SKILL.md     # Claude convention

Node: drop the folders under your working directory. Skills you already have for Claude Code just work.
Browser: seed them via initialFiles at the same paths — no loader call.
Explicit: pass skills: [{ name, description, instructions }] — explicit skills override discovered ones of the same name.

Anatomy of a skill#

.agents/skills/changelog/SKILL.md

---
name: changelog
description: Write a clear CHANGELOG entry from the staged git changes
---

# Changelog Skill
Group changes into Added / Changed / Fixed / Removed, one line each…

A skill can ship a helper script alongside its SKILL.md; when loaded, the agent is told the skill's directory so it can run it.

Why on-demand matters#

Twenty skills would be thousands of tokens if they all sat in the system prompt. The catalog line per skill costs almost nothing, and the model pulls in full instructions only when the task calls for them — so you can be generous about how many skills you ship.

The CLI loads skills from ~/.agents/skills and ~/.claude/skills as well as the project — project skills override home ones, and --skills <dir> adds more.