Skills

Lessons hold facts — "CoinGecko rate-limits at 50 req/min." Skills hold whole procedures — "how to summarize a CSV end-to-end." When you've watched Anton do something well and you want it done the same way next time, you save it as a skill.

Skills live in a single library at ~/.anton/skills/, one directory per skill, shared across all your projects. A skill saved while working in one workspace is available everywhere.

Creating a skill

Do a task, then save it. Anton drafts the procedure from the work it just did — the recent conversation plus the code it ran:

you> Take a quick look at sales_q3.csv

anton> [opens scratchpad, loads pandas, infers schema, prints describe(), plots distributions]
       Here's what I found...

you> /skill save csv summary
anton> Drafting a skill from recent work…
       Saved skill csv_summary → ~/.anton/skills/csv_summary/
       Name: CSV Summary
       When to use: User asks to explore, summarize, or describe a CSV file.

The name hint after /skill save is optional — Anton will pick a sensible label either way. The skill is saved automatically (no interactive editing step) and is available immediately, including for the rest of the current session.

Using a skill

You don't invoke skills manually. Every conversation includes a compact list of your saved skills — just each skill's label and a one-line "when to use" cue. When your request matches one, Anton recalls the full procedure from disk and follows it instead of reasoning from scratch.

Each recall bumps a usage counter, so /skill list shows you which skills actually earn their keep. Anton is also typo-tolerant here: if it asks for a label that doesn't quite exist, the closest match is used.

Skills every artifact build follows

An artifact build can't recall a skill. It follows the skills you mark for it instead. Add applies_to: artifacts to a skill's SKILL.md frontmatter, beside name and description:

---
name: chart-rules
description: How this project's pages draw charts.
applies_to: artifacts
---
- Use only Chart.js for every chart.

The page writers, the backend writer and the step that fetches data then get the skill's instructions. The build follows them and never shows them on the page. Your project's .anton/anton.md reaches every build step the same way, as rules to follow, not text to show.

In the CLI, a build follows the marked skills in your own library and anton's built-in ones. An app that runs Anton names which other skill folders a build may follow. A skill someone else shares with you isn't one of them; it still works when Anton recalls it.

Managing skills

Command	What it does
/skill save <name hint>	Draft a new skill from recent work and save it (name hint optional)
/skill list	Show all saved skills with their "when to use" cues and recall counters
/skill show <label>	Print one skill's full procedure and stats — typo-tolerant, suggests the closest match
/skill remove <label>	Delete a skill from disk

What's inside a skill

Each skill is a small directory at ~/.anton/skills/ named by its label:

~/.anton/skills/csv_summary/
├── meta.json          name, description, when-to-use cue
├── declarative.md     the step-by-step procedure Anton follows
└── stats.json         usage counters

declarative.md is ordinary markdown — open it to read or hand-tune the procedure. The full directory format (including the optional later stages a skill can grow into) is covered in the developer guide at Skills internals.

When to save a skill

Good candidates are tasks that are multi-step, repeatable, and were just done well: a report you'll want monthly, a data-cleaning routine, a deployment checklist. Single facts belong in lessons instead — Anton handles those on its own.