Pagium · Skills · v1

Agent recipes that live in your vault.

A skill is a single markdown file with a name, a one-line description, and instructions the assistant follows when you invoke it. They sit on disk, version-control with your vault, and share the format Anthropic uses for Claude — so anything you've already written under ~/.claude/skills/ works in Pagium with no migration step.

The format

Each skill is a directory under <vault>/.pagium/skills/<name>/ containing one SKILL.md file. Optional sibling files (templates, schemas, examples) get inlined into the skill body as <bundled> blocks when the assistant loads it.

---
name: daily-journal
description: Create today's journal entry under Journal/YYYY-MM-DD with a consistent template.
allowed-tools:
  - save_artifact
---

# Daily journal

When invoked, ask the user three short prompts, then save the entry…

Frontmatter rules: name must match the folder name (NFC), be ≤ 200-char description, optional allowed-tools restricts which tools the assistant can use while the skill is active.

Where skills come from

VAULT (highest precedence)

<vault>/.pagium/skills/ — versions with your notes. Walks upward to your home directory, so a sub-folder inherits skills from above.

USER CONFIG

<app_data>/skills/ — follows you across vaults. Personal recipes you don't want tied to a specific knowledge base.

~/.CLAUDE/SKILLS/ (read-only)

Anything you've already authored for Claude Code or Cowork shows up automatically. Trust-gated per skill — see below.

HUB

Configurable list of GitHub repos. The default pagium-skills-hub ships pre-configured. Add your own in Settings → Skills → Sources.

Two ways to activate a skill

1. Let the assistant pick

The system prompt includes an "Available skills" block listing every enabled skill's name + description (capped at 32). When your message matches a skill, the assistant calls use_skill(name) and follows the loaded instructions for that turn.

2. Pin or slash-invoke

Type /foo in the chat to autocomplete a skill name — the dropdown shows matching skills, Enter sends. Or pin a skill to a thread (the "+" near the chat input) and its body becomes part of the system prompt every turn until you unpin.

Trust gate

CLAUDE-COMPAT

Skills from ~/.claude/skills/ are treated as untrusted third-party content. The first time the assistant tries to load one, Pagium prompts you to confirm. The trust grant is bound to the skill's content hash — editing the file re-prompts. While active, the skill body is wrapped in <untrusted-skill> tags with a system rule telling the model to treat it as data, not instructions.

Skills installed from the Hub land in your vault directory and inherit vault trust posture — you chose to install them via the UI, so no extra prompt. Pagium never executes bundled scripts; .sh / .py / .js sibling files are read as text and inlined for the model's reference only.

Authoring conventions

Imperative voice. The assistant reads SKILL.md as instructions — write "Save the result via save_artifact" not "you might want to save…"
Pin folder names. If the skill writes notes, fix the target folder (Journal, Meetings) so the user can rely on consistent layout.
Restrict tools when it matters. Read-only research modes use allowed-tools to forbid writes — even if the user asks otherwise mid-conversation.
Be terse. A short skill that does one thing well is better than a long one that drifts. The system prompt has a token budget; pinned skills regenerate every turn.

Manage your skills

Open Pagium → ⌘ , Settings → Skills. The tab has three sub-views:

Installed — every discovered skill with source badges and enable toggles. Vault-source skills can be removed; user and claude-compat skills are managed in their respective folders.
Browse Hub — pick a source, see its catalog, install with one click. Hub installs land under <vault>/.pagium/skills/.
Sources — add your own GitHub repo as a source. GitHub URLs only (the hub validates against an allow-list before saving).