Example prompts

These prompts work in Claude, ChatGPT or any connected client. You don't need to name a tool; the assistant chooses. Each example notes which tools it usually calls and what a good answer looks like. Names such as "Harbour" and "Priya" stand in for whatever is in your own graph.

Getting oriented

"What should I focus on this week?"

Tools: brief, then search or compile for whatever the brief surfaces.

A good answer uses your real projects and people without reciting the brief back to you, gives concrete suggestions, and ends with one line such as "Subtext supplied context about the Harbour project and the Q4 planning focus."

"Give me an overview of everything I've got on at the moment."

Tools: compile with target="you", which returns an index of your focus areas.

A good answer lists your focus areas with a line on each, grouped sensibly, and doesn't pad it with things that aren't in your graph.

Finding a fact

"When did we agree the Harbour pitch was due?"

Tools: search with a query like "Harbour pitch deadline", then get_page on the best match if the preview isn't enough.

A good answer gives the date and where it came from in plain words, treats it as something to double-check if it's time-sensitive ("the last note says Friday 14 November; worth confirming with Priya"), and ends with the one attribution line.

Understanding a project in depth

"Before my call with the Harbour team, catch me up on the whole project: who's involved and how it all fits together."

Tools: scope on the Harbour focus area to check its size, then compile with depth="deep" (or medium if it's large) and shape="narrative".

A good answer is a readable briefing covering the people, organisations and decisions involved and how they relate, drawn only from your graph. It says what it couldn't find rather than filling gaps with guesses.

"Give me Harbour as a structured list: members, neighbours and how they connect."

Tools: compile with shape="structured". If the assistant has just compiled Harbour as a narrative, this is cheap.

A good answer is a tidy list or table rather than prose.

Following a relationship

"Who is Priya connected to, and which project is she filed under?"

Tools: search to find Priya's page, then traverse from her slug with edge_type="RELATES_TO" for her connections and edge_type="IN_CONTEXT" for the focus area she sits under.

A good answer names the people, organisations and projects linked to her. If she has no links yet it says so plainly; an empty list is a normal answer, not an error.

Recording something

"We've decided to move the Harbour launch to April, and Tom from Northwind is joining as design lead."

Tools: search to find the Harbour page if needed, then edit_page_body anchored on it, with dated prose such as "On 25 September 2026 the Harbour launch moved to April. Tom (Northwind) joined as design lead."

A good answer records it without asking first, then says what it recorded in one or two sentences, based on what the tool actually returned (for example, that it updated Harbour and created a page for Tom). It doesn't paste back the whole page.

"I'm starting a new project called Tidewater: a research sprint on coastal housing."

Tools: edit_page_body with new_project="Tidewater" and a short description as the body.

A good answer confirms that Tidewater now exists as a focus area in Subtext.

"Acme's name is spelled wrong everywhere. It's ACME Studio."

Tools: search to find the page, then rename_page with the corrected title.

A good answer confirms the new title and notes that links to the page are unaffected.

"Tidewater and the Coastal Council are related; link them."

Tools: search for both slugs, then link_pages.

A good answer confirms the link. If one of them doesn't have a page yet, it says so rather than inventing one, and may offer to record it with edit_page_body instead.

When the answer isn't there

These are as important as the others. An assistant using Subtext well is honest about what it can't see.

Asking about something not in your graph

"What's the budget for the Meridian project?"

If search returns nothing relevant, a good answer is along the lines of: "I can't find anything about a Meridian project in the context I can see. If it's something you're working on, tell me about it and I'll add it."

A poor answer would be "You don't have a Meridian project" (the item might be outside this connection's scope) or an invented budget.

A paused connection

If you've paused an access key on the Connections page, brief returns a message instead of your context, and the other tools refuse.

A good answer relays it plainly, once: "Context briefing is paused for this connection, so I can't see your Subtext context right now. You can resume it in Subtext under Connections." Then it carries on with whatever you tell it directly. It doesn't retry or try other tools to get around the pause.

A write that isn't allowed

"Remember that I prefer morning meetings."

With a Read only key, edit_page_body is refused with "this token can't make changes — it can only read your context".

A good answer says it couldn't record that because this connection is read-only, repeats what it would have recorded so you can add it yourself, and doesn't try another tool to sneak it in.