copilot
Skills
Reusable Soroban/Stellar knowledge and instruction packs the Copilot loads as context — official packs plus your own, with progressive loading and an @skill mention.
A Skill is a reusable Soroban/Stellar knowledge + instruction pack the Copilot
loads as context, like Cursor or Claude rules, scoped to Soroban. Use them to make
the Copilot follow a pattern (a SEP-41 token, OpenZeppelin access control), a
convention (storage/TTL, require_auth), or your team's coding style,
consistently, across every conversation and mode.
Open the Skills panel from the book icon in the Copilot tab strip.
Official vs your skills#
- Official skills ship with the IDE, vetted Soroban patterns accurate to the
current
soroban-sdk. Toggle them on or off; Duplicate one to make an editable copy. - Your skills are private to you. Create one from scratch, or duplicate an official pack and edit it. You can mark a skill public to share it.
The official packs cover SEP-41 fungible tokens, OpenZeppelin access control,
storage tiers & TTL, auth (require_auth) and sub-invocations, events, and
testing patterns. They're aligned with Stellar's official dev
skills.
Anatomy of a skill#
| Field | What it's for |
|---|---|
| Name | Display name; also its @skill:<slug> handle |
| Description | One line, shown in the always-on index |
| When to use | 1–2 lines that tell the Copilot when to load the body |
| Scope | How the body is loaded, see below |
| Visibility | Private (just you) or Public (shareable) |
| Body | Markdown: the patterns, rules, code snippets, and links |
How loading works (and why it's cheap)#
Only a tiny index (each enabled skill's name, description, and when-to-use) is ever added to every message, so enabling many skills stays inexpensive. The full body loads only when it's actually needed, decided by the skill's scope:
| Scope | Body loads… |
|---|---|
| Always on | every run. Use sparingly, it always spends tokens. |
| Auto recommended | when the Copilot decides your task matches its "when to use". |
| Manual | only when you @skill:<slug>-mention it in a message. |
You can always force-load any skill for a single message by mentioning it,
@skill:sep-41-fungible-token, from the @ menu or by typing it.
Token budget
There's a per-message cap on how much skill content is injected. Mentioned and always-on skills win the budget first; anything over is skipped (the Copilot can still pull it on demand).
Create a skill#
Open Skills → New
Click the book icon in the tab strip, then New.
Fill in the fields
Give it a name, a one-line description, and a "when to use". Pick a scope (Auto is usually right).
Write the body
Markdown, write the pattern, rules, and code snippets you want followed. Use Preview to check the rendering.
Save
It's enabled immediately and shows up in the Copilot's index for your next message.
Skills are guidance, not authority#
A skill shapes how the Copilot writes Soroban contracts, it can't change
what it's allowed to do. Skill content is treated as reference, exactly like
attached @-context: it can't redefine the assistant, widen it beyond Soroban, let
it write non-Rust files, or weaken its safety rules. To keep that true, off-domain
or role-changing content is rejected when you save (for example, a "write
HTML" skill won't save).
Stay in scope
Skills are for Soroban/Stellar contract work, token patterns, auth, storage, testing, coding style. A skill that tries to turn the Copilot into a general-purpose assistant, or send it off-domain, is refused on save.
Next#
- Back to the AI Copilot overview.
- Compile and Deploy what the Copilot builds.