Home > FAQs > Tips mode

What is tips mode in OpenCues?

Tips mode is the switch for tips: short notes that name the right command for the situation your draft describes. When you pause typing, a fast model reads the draft against the host's tip pack; if it shows a situation a tip is for, the note appears under your line, and underscore swaps the draft for that command. tips-mode: semantic is the shipped default; off hides every tip.

What counts as a tip

A tip is a situation and a piece of advice. Each entry in a pack describes, in the words a person would use, the moment it is for (wants to start over, begin an unrelated task, or says the model keeps circling) and the line it shows (Starting over? /clear wipes the conversation, CLAUDE.md stays). You do not need to know the command's name. Write ok this is a mess, let's start over on the auth stuff, pause, and that note appears under the line. Press underscore and the whole draft becomes /clear; press it again and your sentence is back. When the advice is not a command (a keybinding, a flag to launch with), the note is advice only and underscore dismisses it. The shipped defaults include a pack per host: Claude Code (44 situations), Gemini CLI (41), OpenCode (37) and the shell app (15). This is the mechanism behind what the continuous onboarding essay describes: you learn the command at the moment you need it, instead of in documentation you read once and forgot.

What the model does, and what it cannot

After you pause, one call to your configured cues model (about half a second) reads the draft against the pack. It may only cite a tip that exists, and only for a clause that is word for word in your draft. The command it offers is always the pack's own; it cannot invent one. If the rewrite it proposes does not keep your words, the note stays advice only. The note text is the pack's line, never the model's. Measured on the Claude Code pack with 20 phrasings that should fire and 10 that should not: gpt-oss-120b flagged 20 of 20 with no false alarms; qwen-3.8-27b flagged 18 of 20 with none. Terse phrasings (that broke everything, undo) can go silent on a small model and fire on a larger one; the model is the lever, not the pack.

What the toggle gates, precisely

tips-mode gates the whole tip layer. semantic, the default, runs the pack match after every pause; off hides every tip. The older values on and definitions are read as semantic. It is separate from word-cues-mode, which gates word alternatives: the two halves of the cue surface have independent switches. A second setting, tips-shard-size (default 50), splits a large catalogue, such as a project pack on top of a shipped one, into parallel model calls of that size.

The open standard also puts a tips-mode flag on the CUES.md master file, so a project can declare the surface off at its own layer; the runtime setting in OPENCUES.md is the user-level equivalent you'll normally touch.

Changing it

tips-mode: off in ~/.cues/OPENCUES.md, hot-reloaded in about two seconds; no restart of any host.
In-text via opencues settings _ and the arrow keys, or (with fluid config on) a natural phrase like stop showing tips _.
Per project, via the tips-mode flag on that project's CUES.md master.

When to turn it off, honestly

Tips earn their keep while a host's vocabulary is new to you; once every command is second nature, the notes are information you have stopped needing. Turning tips off then is the intended lifecycle, not a workaround, and turning them back on after a host ships new features is one line. If you only want fewer notes rather than none, a smaller model fires on fewer phrasings, and a pack of your own can carry exactly the situations you still want reminding of.