Teach Starko a procedure once.It loads it only when it applies.
A skill is an internal article in the Claude skill format: a body that becomes SKILL.md, plus references, scripts and assets. Publish it, and the same version reaches ⌘L, Starko in Slack, your MCP host and Starko AI's coding sessions.
Release checklist
All changes savedSaveRelease checklist
When to use
Use when someone asks to cut, ship or verify a release of the web app.
Before you cut
- Check that the main branch is green and nothing is waiting for review.
- Run scripts/check-release.sh with the version number and report what it prints.
- List the merged pull requests since the last tag.
If something breaks
Follow references/rollback.md and post in the release channel before you roll back.
Files
Publishing
Available in
MCP: shows up as a prompt in Claude, Cursor and VS Code (not for "Only Starko" skills).
Who can use it
Projects
Which projects' coding sessions get this skill.
Scripts
Versions
Skills in short.
- Format
- The Claude skill format:
SKILL.mdplusreferences/,scripts/andassets/. - Available in
- ⌘L chat, Slack, MCP and Coding sessions. New skills start on everything except MCP.
- Who can use it
- People and Starko, Only people via /, or Only Starko.
- Versions
- Publishing freezes a version. Agents only ever read published versions.
- Scripts
.sh,.pyand.js, run in an ephemeral sandbox after you approve, 60 seconds per run.- Network
- Off unless the version lists domains, up to 10. Only owners and admins set them.
- Limits
- SKILL.md up to 20,000 characters, 50 files and 8 MB per skill.
- Visibility
- Always internal. A skill never appears on your help center or to customers.
A skill is an article with a name and a line that says when to use it.
Skills live in the Knowledge hub, on a Skills tab next to Articles. New skill opens "Teach Starko a skill": start blank or from an internal article, give it a title, a name and When to use. When to use is the part Starko reads to decide whether to load the skill, so it says when, not what. The body is written in the same editor as any article.
- The name becomes the skill's folder: lowercase letters, digits and hyphens, up to 64 characters. It is fixed once a version is published.
- When to use can be up to 1,024 characters.
- Any internal article's ⋯ menu has Turn into a skill. A skill leaves knowledge base search, because Starko loads it whole instead of quoting passages from it.
- The list shows each skill's name, When to use, where it is available, its live version and when it changed.
| Name | When to use | Available in | Status | Updated |
|---|---|---|---|---|
| Release checklistrelease-checklist | Use when someone asks to cut, ship or verify a release of the web app. | ⌘L chatSlackCoding sessions | v3 | Oct 28 |
| Refund decisionrefund-decision | Use when a customer asks for a refund and someone needs to decide whether it qualifies. | ⌘L chatSlack | v2 | Oct 21 |
| Weekly support digestweekly-support-digest | Use when someone asks for a summary of last week's conversations and tickets. | ⌘L chatSlackMCP | v1 | Oct 19 |
| Web app code styleweb-app-code-style | Background for code changes in the web app: naming, tests and folder layout. | Coding sessions | v5 | Oct 12 |
| Incident write-upincident-write-up | Use when someone asks for the summary of an outage after it is resolved. | ⌘L chat | Not published | Oct 30 |
Bundle the references, scripts and assets the procedure needs.
The Files card under the body is the skill's folder. Upload files or add one by hand, and each lands where Claude expects it: scripts in scripts/, Markdown and text in references/, everything else in assets/. Text files open for editing in place. At the top sits SKILL.md, generated from the name, When to use, who can use it and the body. Its preview is exactly what agents receive.
- Refer to a file from the body by its path, for example
references/rollback.md. Starko reads it only when the instructions point to it. - Who can use it becomes frontmatter: Only people adds
disable-model-invocation: true, Only Starko addsuser-invocable: false. - Only owners and admins can add or edit scripts. Members see them locked.
- Publishing refuses a SKILL.md over 20,000 characters, so agents never get a cut-off procedure. Long material belongs in references.
SKILL.md
What agents receive: generated from the name, When to use, who can use it and the body.
--- name: release-checklist description: "Use when someone asks to cut, ship or verify a release of the web app." --- ## Before you cut 1. Check that the main branch is green and nothing is waiting for review. 2. Run `scripts/check-release.sh` with the version number and report what it prints. 3. List the merged pull requests since the last tag. ## If something breaks Follow `references/rollback.md` and post in the release channel before you roll back.
Drafts stay drafts until you publish a version.
Editing a skill changes nothing Starko runs. Publish creates an immutable version, and every agent reads the live version only. The Publishing section shows the live version and date, whether the draft has unpublished changes, and the button for the next version. Publishing a draft that matches the live version creates nothing new.
- Body, When to use, who can use it and network domains are draft content and take effect on publish.
- Available in and Projects are distribution settings and apply at once.
- A skill with scripts or network access can only be published by an owner or admin.
- Versions lists who published each version and when, with the live one marked.
Starko picks a skill when a request fits, or you call one with /.
In ⌘L and Starko in Slack, Starko sees the name and When to use of every published skill it may pick, and loads one when a request matches. To call one yourself, type / in the ⌘L composer for a menu of the skills you may use, or write $name. A skill you invoke shows as a chip in the sent message, with its version, its description, View instructions and Copy.
- Each row in the menu has a Workspace badge, or the projects the skill is limited to.
- Money like $20 never opens the menu or counts as a skill.
- Up to 3 skills you name in one message are loaded with it. Only Starko skills cannot be called by name.
- Skills are for your team. Public and embedded chat sessions never get them.
Release checklist
Version 3Use when someone asks to cut, ship or verify a release of the web app.
Bundled scripts run in a sandbox, and only after you approve.
When a skill's instructions say to run one of its scripts, Starko asks first. The approval names the file, the skill and its version, the arguments and the network access. Approve, and the script runs in an ephemeral sandbox with no secrets and no environment variables, which is deleted afterwards. The result card shows the file, the duration, the exit code and the output.
.shruns with bash,.pywith python3,.jswith node. Up to 20 arguments, passed exactly as approved.- A run stops after 60 seconds, and Starko gets the last 8,000 characters of output.
- Network is blocked unless the published version lists domains. Then only those domains are reachable.
- If the skill is republished between your approval and the run, nothing runs.
- Every run asks, in ⌘L and in Slack, including scripts with no network access.
version 2.14.0 is not tagged yet changelog entry found migrations: 2 new, both reversible ready to tag
The check passed with exit 0. Version 2.14.0 is not tagged yet, the changelog entry is there and both new migrations can be rolled back. Next step on the list: the merged pull requests since 2.13.2.
The same skill in Claude, Cursor and VS Code, and in Starko AI's checkout.
Switch on MCP and the skill appears as a prompt in your host's own / menu through the Starko MCP server. Picking it inserts its SKILL.md, plus your request if you add one. Switch on Coding sessions and the published version goes to Starko AI's fix runs and pairing sessions, for the projects you pick or all of them.
- Only Starko skills stay out of the MCP menu. Reconnect the connector or start a new conversation to see newly published skills.
- In a coding session the agent finds skills in a
skills/folder in the checkout, kept out of git so it never lands in a pull request. - The Publishing section shows when a version is ready for coding sessions, or still uploading.
- Sandbox setup, project learning and evaluation replays run without skills, so their output stays your project's own.
- A coding session gets at most 50 skills.
FAQAbout Skills.
Connect a channel and a repository. Turn the support queue into a build queue.
Feedback starts flowing in minutes. The first recommendation shows up as soon as a pattern repeats, and the first pull request as soon as you accept one.