BuildingSkills

Skills

A skill is authored knowledge — markdown plus attachments — that lives in a workspace's or a member's personal Context library and gets packaged into an Agent's compiled artifact when equipped, the same way an MCP connection is. Where a connection gives an Agent tools, a skill gives it something to know — attach one from the Context section of the agent editor.

The skill editor

A skill has a Name, a Description ("A one-liner the agent uses to decide when to reach for this skill"), and an Instructions body — the same rich-text/markdown-toggle editor used for Persona, with the placeholder "# How to… Write the skill in Markdown." An attachments sidebar sits alongside the body.

Saving is explicit, not autosaved — a real behavioral break from the agent and workflow editors, both of which save every keystroke to the draft. The header tracks dirty state instead: "Unsaved changes" while you're editing, a green "✓ Saved" once a save lands. Nothing you write persists until you press Save.

A read-only member — anyone without manage rights on a workspace-scope skill — sees the identical layout with every mutating control removed.

Attachments

Drop files onto the sidebar, or click "browse." Accepted formats are text: .md .markdown .txt .json .jsonc .yaml .yml .toml .csv .tsv .xml .html .css .js .jsx .ts .tsx .py .sh .rb .go .rs .java .sql .env .ini .log, plus anything the browser reports as text/*. The server enforces this independently — a binary upload is rejected with a 415 regardless of what the client sent. Re-uploading a name that's already attached replaces the existing file.

Limits

WhatLimitOn exceeding it
Instructions body256 KiB of markdownThe save is rejected
Each attachment5 MiBRejected client-side per file with a toast — "filename is larger than 5 MiB" — the rest of the batch still attempts to upload

See limits & defaults for the platform's other ceilings.

Skills, personas, and connections

Three different things equip an Agent, and it's easy to reach for the wrong one:

What it's forWhere it livesChanging it
PersonaOne Agent's identity and working styleOn the Agent itselfAutosaves to that Agent's draft only
SkillReusable knowledge, shared across any Agent that equips itThe workspace or personal Context libraryExplicit save; affects every Agent that has it equipped, on their next publish
ConnectionTools — an MCP server's capabilitiesThe workspace or personal Context libraryManaged in Context & MCP; affects every Agent that has it equipped, on their next publish

Deleting a skill

Deleting a skill removes it and its attachments outright — unlike a connection, it isn't blocked while something still references it. Nothing fails immediately: an Agent that still has it attached reports it as missing in the agent editor's Context section, and can't publish again until the reference is removed or replaced — but every already-published version keeps the copy that was baked in at its publish. Prose that merely names the skill — a persona, a workflow step's instructions — is never checked against the library, so nothing flags it; it just goes quietly out of date.

Changes land on the next publish

Editing a skill doesn't retroactively touch an Agent that's already published — a published version is an immutable, content-hash-pinned artifact, and the skill's content at publish time is what got baked in. Edit the skill, and every Agent equipped with it picks up the change the next time each one publishes, not before.