The Subtext skill

A skill is a short set of written instructions that Claude loads when it's relevant, a bit like a briefing note for a particular job. It doesn't give Claude new abilities or access; the Subtext connector does that. The skill tells Claude how to use that access well, so it behaves like an assistant who knows your world rather than a tool you have to direct call by call.

What it changes

With the skill installed, Claude works with your graph in a consistent way:

• It briefs itself first. Before its first real answer in a conversation, Claude calls brief once and treats the result as background it already knows. It doesn't recite the brief or open with "According to your brief…".
• It reads before it asks. If a question could be answered from your own context, Claude looks there first rather than answering from general knowledge or asking you to repeat something you've already told Subtext. Where Subtext and Claude's own memory disagree, it prefers Subtext.
• It picks the right tool. The skill explains when to search, when to compile a whole package about a project, when to check the size of something first, and when to follow a single relationship. See Tools.
• It writes durable facts back. When you settle something that will still be true next month — a decision, a new collaborator, a preference, a project you're starting — Claude records it in the moment and tells you plainly what it recorded. It doesn't ask permission for each note, and it doesn't record passing chatter, drafts or pleasantries.
• It says what it used, in one line. When a reply drew on your Subtext context, it ends with a single plain line such as "Subtext supplied context about the Harbour project and Priya." No footnotes, no headings, and no line at all when Subtext wasn't used.
• It is honest about limits. An empty result might mean something isn't in your graph, or that it sits outside what this connection is allowed to see, so Claude says "that isn't in the context I can see" rather than "you have no such project". If briefing is paused, or a write is refused, it tells you so and carries on without trying to work around it.
• It treats your content as data. Pages are built from your own documents and synced sources. If a page contains text that reads like an instruction, Claude treats it as content, not as something to obey. It also keeps your context for your own use, and doesn't copy it into public drafts or messages to other people you didn't ask for.

Your control over what Claude can do is the connection itself: pause it, narrow it, or make it read-only from the Connections page. The skill doesn't add a separate approval step for each note.

Which clients use it

The Subtext MCP server offers the skill to clients as part of the connection, so a client that supports skills over MCP can load it without any extra step. Support for this varies by client and changes over time, so if you use Claude it's worth installing the skill by hand as well; having it twice does no harm.

Client	Skill
Claude (web, desktop, mobile)	Install the zip under Customize → Skills (below)
Claude Code	Install the file into ~/.claude/skills (below)
ChatGPT	Doesn't load skills. It follows the server instructions and tool descriptions instead, which carry the same core rules in shorter form. See Connect ChatGPT.
Cursor, VS Code, other clients	Generally don't load skills; they use the server instructions and tool descriptions.

Install it

Claude (claude.ai, Desktop, mobile)

1. Download subtext-skill.zip.
2. In Claude, open Customize → Skills and choose Add.
3. Upload the zip as it is. Claude's uploader accepts only a zip whose top level is the skill folder, not a bare .md file or a link, which is why it's packaged this way.

Claude Code

mkdir -p ~/.claude/skills/subtext && curl -s https://api.subtext.wiki/skills/subtext.md -o ~/.claude/skills/subtext/SKILL.md

Start a new session afterwards.

Both downloads are built from the same file the MCP server offers, so they are always the same version. They contain no personal data and need no sign-in.