Skip to main content
Context GrammarDesign for context-aware AI

Context Grammar · Stage 3

Relationship Dials: control without friction

Know me. Help me. But let me choose what you share and what you do.

Relationship Dials

English audio

Read transcript

The usual—but with care

At your regular restaurant, ‘The usual’ can be enough. The server remembers what you like, and that feels wonderful. But if today's special costs much more, you would want them to ask first. And if they adjust a dish for your dietary needs, you would not want those needs announced to the friend beside you. Good service has two boundaries: what may be shared, and what may be done without asking.

Two different decisions

AI needs the same care. The Disclosure Dial sets what an assistant may know, reveal, and pass along. The Autonomy Dial sets what it may actually do for you. Knowing your preferences is not permission to buy something. Helping with a private need is not permission to tell everyone about it. These are two separate decisions, adjustable for each person and each task.

Disclosure · Protect

First, Protect: how much of your information can be used for this task? Imagine checking a calendar to arrange a meeting. Full shows the event and names. Summary shows only a broad time window. Existence says merely that the time is taken. Hidden reveals nothing. Watch the details disappear, one layer at a time. The useful level depends on the job and on who is asking.

Disclosure · Connect

Next, Connect: even when sharing is allowed, when should it arrive? Your mother watches a film she loves and wants to recommend it. Push interrupts you now. Digest brings the recommendation with other family notes this weekend. Available waits until you ask. Off sends nothing. The same warm thought can feel welcome or intrusive, depending on its timing.

Disclosure · Sync

Sometimes two people each believe a different plan. You remember Saturday at the east gate; your friend remembers Sunday at the west gate. If both of you allow it, your assistants can compare only this appointment and flag the mismatch. They do not exchange your whole calendars. And they do not choose the new plan for you. You and your friend settle it together.

Deep understanding. Careful delivery.

These choices work together without becoming one setting. Your assistant may remember everything you said about a film. Your family digest might share only its title and a short invitation. Understanding you deeply does not mean broadcasting every detail. Protect controls the depth of access. Connect controls the delivery. Sync coordinates a permitted exchange across assistants.

Now: the Autonomy Dial

So far, we have been talking about information. Now ask a different question: how far may the assistant act? The Autonomy Dial has four levels: Suggest, Confirm, Notify, and Auto. Let's follow one family from a new shoe purchase to an everyday household routine. Notice that the level can change with the task.

01 · Suggest

At Suggest, the assistant finds running shoes that fit the child's size, tastes, and deadline. It shows a short list, then stops. The parent compares, chooses, and starts any purchase themselves. This is guidance, with no order prepared or placed.

02 · Confirm

At Confirm, the assistant can prepare more: the right size, delivery date, address, and a checkout draft. But the order does not go through until the parent reviews and approves it. The distinction is visible: prepared is not purchased.

03 · Notify

For a familiar household item, you might give narrower permission. At Notify, the assistant orders the usual coffee beans within the budget and brand you set, then tells you what it did. The receipt and an easy way to stop the routine remain available.

04 · Auto—and an exception

At Auto, that predictable refill can happen quietly, without a notification every time. But Auto is not a blank cheque. If the usual beans are sold out and the substitute exceeds your price limit, the assistant pauses and asks. A boundary is most meaningful when something unexpected happens.

One household. Different boundaries.

The assistant may know a shopping budget without seeing your whole banking history. It may know the child's shoe size without being allowed to buy shoes. In a shared home, routine supplies can be automatic while a child's new activity waits for a family decision. Permission belongs to a person, a task, and an audience. One family member's yes is not everybody's yes.

Shape how help arrives

The same design reaches everyday recommendations. Your taste can guide what is found. Connect can save discoveries for the weekend instead of interrupting work. Autonomy decides whether the assistant merely suggests a song or starts playing it under a permission you chose. You can ask for more surprises today, and more quiet tomorrow.

Pause is not erase

And changing your mind should be easy. Pausing future music sharing stops new updates. It does not pull back messages already delivered. Deleting remembered preferences is another, separate choice. A clear interface shows what has stopped, what remains, and how to resume. Trust grows when you can correct a boundary without losing your place.

Help, on your terms

Back at the restaurant: ‘The usual, please. But I'm here with a friend, so let's keep it quiet.’ The server understands both the request and its limits. That is the relationship worth designing with AI: enough understanding to help, clear permission to act, and care about what stays private. Disclosure and Autonomy let you adjust that relationship, one intention at a time. Explore the chapter for the full framework, and continue to the Rule Engine to see how these choices become behavior.

A host leans in to speak quietly to a regular while her friend reads the menu.
“The usual” — confirmed quietly. What may be shared and what may be done are two separate boundaries.

At the table

When care crosses a line

At a restaurant you know, “the usual, please” can be a relief. The server remembers what you like and brings a familiar dish without another round of questions.

But “the usual” is not permission to bring an expensive tasting menu. And knowing about a private dietary need is not permission to announce it to your guest. The same attentiveness that feels thoughtful can feel intrusive when it crosses a boundary.

A personal AI needs two separate answers: what may it learn or reveal, and what may it actually do? Context Grammar calls these the Disclosure Dial and the Autonomy Dial. You can change each answer for each person and each kind of task.

A familiar host warmly welcomes a regular at the door
The comfort of being remembered; the safety of being asked.
01

Remember

The server knows your usual order.

02

Ask

An unfamiliar, expensive choice needs your answer.

03

Be discreet

A private detail stays between you and the server.

Two independent questions

Knowing is not permission to act

You might let an assistant know your child's shoe size, your budget, and that PE is tomorrow. That helps it find suitable shoes. It does not authorize a purchase. You can share rich context while keeping the final tap entirely yours.

01 / Disclosure

What may this assistant read, and what may it tell this particular person?

02 / Autonomy

May it suggest, prepare, act and report, or run a preapproved routine?

Two rules

How the two dials work together

The dials are set separately, but they are not unrelated. Two rules connect them.

a

Enough, and no more

To act, the AI uses only the context that this action needs and that it is permitted to use. More data is never a reason to do more.

b

The more it may do, the more it shows

The wider the autonomy, the more the AI must show what it did. A suggestion needs little explanation; a routine that runs on its own keeps a log you can read. The AI can't lower that duty itself.

Disclosure · Protect

Person → AI: what it may know

Disclosure runs in three directions. Protect is the first: from you to the AI. It sets what this assistant may know about you, area by area. Four levels make the difference visible.

01 / 04

Full

Permitted details

Your shopping assistant may know the size, budget, and delivery deadline.

02 / 04

Summary

A useful abstraction

Your travel assistant knows to avoid long walks, not the medical reason.

03 / 04

Existence

Only that something exists

Your scheduling assistant knows Tuesday evening is taken, not what it is.

04 / 04

Hidden

No view

A private medical note never enters this conversation.

These levels are set per area and per assistant, not once for your entire life. In a family, also ask who may decide. A parent can choose what the AI knows about the family budget, but being a parent is not automatically permission to have the AI watch a child.

Disclosure · Connect

AI → people: what it tells, and when

Connect is the second direction: from the AI to people. It decides what each person is told and when. Give a person as little as changes a decision: a friend hears that Sunday afternoon is difficult, not why. Then choose the rhythm, because a useful reminder can still interrupt at the wrong moment.

Push

Now

A school pickup change that needs an immediate decision.

Digest

At a calm interval

A short evening roundup of non-urgent family updates.

Available

When asked

A restaurant suggestion stays ready without becoming a notification.

Off

No transmission

This topic produces no updates for this recipient.

Disclosure · Sync

AI ↔ other AIs and services: what passes on a handoff

Sync is the third direction: between your AI and other AIs or services. You and a friend both think you agreed to meet, but one calendar says Saturday and the other says Sunday. With permission from both sides, your assistants can compare only the meeting details needed to notice the mismatch. Neither needs the other's complete calendar.

They can ask, “I have Saturday at 2; you have Sunday at 2. Which is right?” The people settle it. Sync is a pattern for limited, permissioned coordination—not a claim that agents may exchange everything they know. A future implementation could use targeted attribute comparison or stronger privacy-preserving methods; the essential rule is to reveal only the needed difference.

Telling a person and briefing another agent are different depths of the same state. A person gets as little as changes a decision. An agent gets what its task needs. The person's private notes never pass. When one coding session hands off to the next, the next one receives the branch, the open question, and the agreements it must keep—not the notes the person wrote for themselves.

Your calendarSat · 2 pm
Friend's calendarSun · 2 pm
Only the mismatch is surfaced

Saturday or Sunday?

Autonomy Dial

Choose how far the assistant may go

The Autonomy Dial sets action authority for a particular task. Moving upward does not mean the person cares less; it means a specific routine has earned a carefully bounded form of trust. The AI may propose a higher level, but only the person approves it. Moving down needs no approval: when something goes wrong, the level drops automatically.

01 / 04

Suggest

Show me options

It finds eight shoes that fit. You choose and buy.

02 / 04

Confirm

Prepare, then wait

It fills an order with the right size, price, and delivery date. You approve before anything is placed.

03 / 04

Notify

Act within limits, then tell me

It reorders a preapproved household essential within your price cap and sends a receipt.

04 / 04

Auto

Run a bounded routine

It performs an explicitly authorized, low-risk repeat task and keeps a reviewable log. Exceptions stop the routine.

A new seller, a price above your cap, missing stock information, or an irreversible commitment sends the task back for review. Auto is not a blanket license to buy, book, or message on your behalf. Every permission has bounds: scope, purpose, duration (today only, this trip only), recipient, and a ceiling (up to $X). Building with AI reads the same way: it may read and edit code on a branch, but deploying asks first.

Try the dials

Two choices, one relationship

Try a view and an action level for a shoe-shopping assistant. The dials are independent: more personal knowledge never raises action authority by itself.

Information view
Action authority
Full

The assistant can use the permitted size, preferences, budget, and deadline.

Suggest

It may recommend. A person makes the choice and completes the order.

A dial preview is not consent. Every action still needs a task-specific agreement, current facts, and a safe way to stop.

In real life

Make the boundary easy to see and change

01

Scope the agreement

A family AI is not one family permission. Set access and action separately for each person, task, recipient, and device.

02

Keep context separate from delegation

Knowing the exact shoe size improves a recommendation. It does not move Suggest to Confirm, Notify, or Auto.

03

Offer two different ways back

Pause future sharing to close the tap. Delete stored memory to empty the reservoir. Previously delivered information may not be retractable.

04

Make mistakes repairable

Show what happened, make corrections easy, and lower authority quickly when a routine goes wrong.

The next chapter

How do these boundaries become behavior?

Intent describes what the person is trying to do. Signals describe the moment. The Relationship Dials define what the assistant may know, share, and do. The Rule Engine turns those inputs into a response the person can inspect and steer.