How to use it

Two steps to finding out what something does.

1 Ask, in the words you already have.

what can you do about engineering Ask "how does the checklist work" — "what is the gap page" It answers from these pages, so the app and the site cannot disagree.

2 Or press the question mark on any card.

Any settings card Push-to-talk Right Shift ? opens this page at that row The mark takes you to the section for that exact row, not to the top of a page.

! The one that stops people.

Help is about the page, not the subject. It explains D47, not Elite. For what a Guardian beacon does, ask the model instead.
Why it works this way

How to ask what something does, and where the answer comes from.

1 Two ways to ask, and both mean “this page”.

? press the mark top right of the panel or just say so “help” · “what is this” no model needed, no network Both open help for whatever you are looking at. Saying it matters more than pressing it: with a headset on, your hands are on the stick and there is nothing to point at the mark with.

2 It sits on top, and “back” puts it away.

the page right where you left it help back help, over the top of it nothing underneath moves “Back” is the same word that leaves anything else. Say it, press the breadcrumb, or use the controller button — all three are one gesture.

3 The pictures are here. The detail is on the site.

IN HERE the pictures, in your theme offline, and inside the headset ON THE SITE the tables, the settings, the reasoning and the working Every page in here ends with a way through to its long form. Not every page has pictures yet. The list below is the ones that do.

And a page nobody has illustrated yet still has its mark: pressing it, or saying so, brings you here instead — one level broader than you asked for, rather than nothing at all.

The details

What Directive 47 can actually do, answered from its own list of capabilities rather than from the model’s memory.

Ask for it

“what can you do” “what are your capabilities” “tell me about the voice capabilities”

This one works with no model configured and no network — it never needs to ask anything outside the machine what is on the machine.

What you get

An overview first, because reading forty capabilities aloud is not help:

I have 9 capabilities, in these groups:
  Foundation — Journal, Diagnostics, Privacy, Settings
  Conversation — Conversation
  Voice — Speech, Callouts
  Interface — Interface
Ask about a group by name for the detail.

Then, by group, the detail and the phrases that actually work — your words rather than internal names, because a list you cannot say out loud is a list you cannot use:

Speech: Speak replies aloud, mark each loop state with its own cue, and stop on command.
  Try: "stop"; "be quiet"; "shut up"
Callouts: Speak up about danger, fuel, route progress and arrivals without waiting to be asked.
  Try: "what are you watching for"; "stop calling things out"; "start calling things out"

Ask for a group that does not exist and it names the real ones instead of refusing. If you asked for the wrong group you want the right one, and this is the moment Directive 47 knows both:

I have no group called "Navigation". I have: Foundation, Conversation, Voice, Interface.

Why it will not make things up

The model is never asked what Directive 47 can do. Ask a model to describe its own abilities and it produces a fluent, confident, partly invented list — and you have no way to tell which half is which until you ask for something that turns out not to exist.

So the answer is built from the actual list of registered capabilities. It cannot name one that does not exist, because it is assembled from the ones that do.

The order it reads them in

Within a group, whichever you have used most comes first, so the things you actually reach for rise to the top of the list. Groups themselves stay in a fixed order — a spoken list whose headings move between askings is harder to follow than one that does not.

Reaching for something counts even when it fails, since a capability you keep trying is worth hearing about early on the days it is not working.

The count covers the current session only. It starts fresh each time Directive 47 does, rather than claiming a usage history it does not have.

The tool surface, for contributors

get_capabilities

Projects the registry: groups, the capabilities in them, and their declared example phrasings. The optional group name expands one of them.

{"type":"object","properties":{"group":{"type":"string","description":"Optional group name to expand, such as Voice. Omit for the overview."}},"required":[],"additionalProperties":false}

The keyword router only routes to a tool with no required parameters, because it deliberately does not extract argument values from free text — a router that guesses at arguments is one that calls the right tool with the wrong ones. The router originally refused any tool with parameters at all, which made “what can you do” match nothing: the one capability whose entire purpose is being answerable without the model was unreachable without it, purely for offering a refinement it does not need. The rule is now “no required parameters”, and the router invokes with empty arguments.