Skills

A skill is a saved playbook Lucius can load into a conversation when relevant. Teach him a workflow once (“here’s exactly how I want a monthly close done”) instead of re-explaining it every time.

Creating one

From the Skills tab, + New skill offers four ways in:

Option	What it does
Create with Lucius	Chat with him to design the skill together
Write skill instructions	Author the name, body, and tool list yourself
Upload a skill	Install a .zip bundle (from Claude Code, or shared by someone else)
Browse starter skills	Pick from built-in templates and adjust

A skill Lucius creates or edits on your behalf goes through the same confirm-style review as any other write before it’s saved.

What a skill actually is

A folder whose only required file is SKILL.md: a short YAML header (name, description required) plus plain-Markdown instructions. Open Net Worth uses the same open Agent Skills folder format Claude Code uses, so a skill from Claude Code mostly works here as-is, and one you build here exports the same way.

my-skill/
├── SKILL.md
├── scripts/        # optional helper scripts
├── references/      # optional docs Lucius reads on demand
└── assets/           # optional static files

The description is what Lucius sees at all times, for every enabled skill — enough to know when it applies, not the full instructions. He loads the full body only when he picks the skill.

How a skill gets used

• Lucius picks it himself. Every turn, he sees name + description of every enabled skill; a match loads the full instructions.
• You pick it explicitly. Type / for a picker, or /skill-name directly — deterministic, for when you want a specific playbook to run for certain. Stack up to six in one message, with trailing text as an argument (e.g. /monthly-close July).

Once loaded, a skill stays loaded for that conversation

No “unload” action — no reliable way to erase instructions a model has already read. Want a clean slate? Start a new conversation. Deliberate: the app won’t show a “deactivate” control implying it can guarantee something it can’t.

What a skill can’t do

It cannot grant, hide, or pre-approve tools. Its instructions can tell Lucius what to do and in what order, but every tool call is still subject to the exact same permission level that tool has outside the skill. “Just always send the email without asking” doesn’t override a tool set to needs_approval — the permission ladder is the one authority on whether a call fires.

Running scripts from a skill

An executable script goes through the sandbox’s own kill switches, including an “unattended script runs” gate that defaults off — so a scheduled Automation can’t run a skill’s scripts headlessly until you enable that deliberately in Settings → Developer. On top of, not instead of, the ordinary permission ladder.