Skip to main content
You can leave notes for the agent inside the document itself. Wrap them in any of these delimiters:
  • [[ ... ]]
  • (( ... ))
  • << ... >>
Pick whichever doesn’t collide with the syntax of what you’re writing (LaTeX uses [[ ]] for some macros, for example, so (( )) works better there). The agent recognizes all three.
A document with inline [[ ... ]] directives left for the agent
You keep writing. The agent auto-wakes three seconds after you stop typing (see Auto-wake on idle), reads the open tabs, pulls out every [[ ... ]] it finds, treats each one as a task in a queue, and handles them (proposing edits, posting comments, or running tools as appropriate). When it acts on a directive, it deletes the directive text from the document so the prose closes back up. If you don’t want to wait the three seconds, Cmd/Ctrl+Enter in the editor triggers the same wake-up immediately. Clicking the agent pill at the top of the history pane does the same.

​
Why this exists

Talking to the agent through the chat popover is great when you have a clear, immediate request. But the requests that actually come up while writing are often local and contingent: “fact-check this,” “I’m not sure about this transition,” “rewrite this once I figure out the rest of the section.” Stopping to type those into a chat box pulls you out of the flow. Directives are a way to write the request where the request applies (in the document, at the spot it matters), and pick it up later, in batch.

​
How they’re handled

The agent’s behavior depends on the autonomy level:
  • Conservative. The agent only edits when a directive is present or the user has made a specific request. Directives are the main way to get edits at this level.
  • Balanced. Directives are handled first; the agent may also pick up obvious problems in the prose.
  • Aggressive. Directives are still handled, but the agent is also proactive about improvements you didn’t write a directive for.
If the agent can’t handle a directive (it asks for a tool you don’t have, or for information it can’t get), it leaves a comment explaining what it would need, instead of silently failing.

​
A few patterns

You can chain them. The agent works through the active tab’s directives first, then any directives on other open tabs.

​
Use them with hooks

Because handling a directive counts as an Edit tool call, your hooks fire. If you have pdflatex wired up, the PDF rebuilds. If you have an auto-commit hook on Stop, the change pushes to git when the agent finishes the turn. The practical workflow: write your draft with directives sprinkled through it, walk away, come back, wake the agent, and watch a batch of small concrete changes land, each one already built and committed by the time you start reviewing.