Skills.

Skills are reusable capability packages from the open Agent Skills standard — the same format used by Claude Code, Cursor, Copilot, Gemini CLI, OpenCode, and other compatible agents. A skill is a folder with a SKILL.md: frontmatter metadata plus markdown instructions. Because the format is shared, skills move in both directions: drop a community skill into your vault, hand your vault skill to any other agent.

Enable them under Settings → Curtis.ai → Skills.

Getting skills in

Three ways, all ending in the same vault folder:

1. One-click import (desktop) — Import from Claude scans ~/.claude/skills, ~/.agents/skills, and the vault's .claude/skills, and copies every skill folder into AI/Skills/. See IMPORT.md.
2. Copy the folder in yourself — any skills-compatible agent's folder works; drop it under AI/Skills/.
3. Write one — Settings → Skills → Open / create makes a starter to edit.

Because the format is the open standard, skills also move out the same way: take any AI/Skills/<name>/ folder and hand it to Claude Code, Cursor, or Copilot.

Format

AI/Skills/
└── pdf-processing/
    ├── SKILL.md          # required: metadata + instructions
    ├── references/       # optional: docs the agent reads on demand
    ├── assets/           # optional: templates, data files
    └── scripts/          # optional: code (run it via run_command if enabled)

SKILL.md is YAML frontmatter + markdown body:

---
name: pdf-processing
description: Extract PDF text, fill forms, merge files. Use when handling PDFs.
---

# Instructions

1. ...

Spec rules the loader enforces: name (1–64 chars, lowercase letters/digits/single hyphens) must match the folder name; description (1–1024 chars) says what the skill does and when to use it — that line is what the model reads when deciding to load the skill. Unknown frontmatter fields are tolerated, never rejected, so future spec revisions keep working. Files that fail validation show up in Settings → Skills with the reason and are never injected.

How the model gets them

Progressive disclosure, in three stages:

1. Discovery — every enabled skill's name + description rides the system prompt as a compact index (~100 tokens per skill) in chats, arena runs, headless runs, and swarm followers. All of them, on every provider.
2. Activation — the full instructions load only when needed:
   • The model calls the load_skill tool (agent mode) when a task matches a description.
   • You attach it explicitly with /skill <name> — works in plain mode with no tools at all.
   • A caller pre-activates it through the plugin API (chat/runAgent skills option) — scheduled runs and MCP clients included.
3. Execution — bundled references/ and assets/ stay unread until a loaded skill points at them; the agent fetches them with the ordinary note tools using the real vault paths.

A per-skill override (Settings → Skills) can force Always in context instead — the full body rides every request. Use it for skills you want applied unconditionally; leave the rest on demand so context stays lean.

Where skills apply

Everywhere inference runs: normal chats, arena mode, swarm leaders and followers, scheduled runs, the MCP server's run_agent, and the plugin API. Agents can pin skills (skillIds on the agent) so a specialist always runs with its procedure loaded. Deliberate exceptions: autocomplete, selection actions, recap, and memory extraction keep their minimal prompts and never see skills.

Controls

Control	Where
Master toggle	Settings → Skills → Enable skills
Skills folder	Settings → Skills (default AI/Skills/)
Starter skill	Settings → Skills → Open / create
One-click import (desktop)	Import from Claude — see IMPORT.md
Per-skill enable + load mode	Settings → Skills rows
Pin to an agent	Agent editor → Skills — rides every run of that agent, swarm followers included; advertised on the swarm roster
Leader attaches per subtask	spawn_agent's skills parameter
Attach to this chat	/skill <name> (/skill bare = list, /skill off = detach all)
Pre-activate headlessly	api.chat(prompt, { skills: [...] }), api.runAgent(name, task, { skills: [...] })

load_skill counts as a vault-class tool: an agent with vault tools off can't pull skill bodies either.

Authoring tips

• Write the description for a model, not a human: say what it does and the trigger — "Use when the user asks to …".
• Keep SKILL.md under ~500 lines; move depth into references/ files and point at them.
• To duplicate a skill, rename the folder and the name field together.
• Skills are instruction content, same trust class as the rest of your vault — the injection fence still applies to note contents and tool output, and skill text is user-authored by definition.