Prerequisites
An actor is doing work the system can see, and there is evidence of typical use: what most actors need most of the time, judged from the task's shape, from general data, or from observed use. Every level of help is a bet on that evidence. It is also settled that the actor leads and the system steps in selectively. These judgements happen before this sequence, and no pattern here records them.
An actor is doing work the system can see, and the system could do part of it. Each level of help takes on more of the work, and a surface moves up or down over its life rather than settling on a level once.
Filling in what is known
Set the default guess — Good defaults
Fill fields with the system's best guess from the task's shape and general data, without asking the actor for anything; the guess also shows the expected kind of answer. Decide which fields get one. Leave out identity, consent, and other sensitive choices, and use a guess only where it is acceptable for it to go through unread.
Replace the guess with known values
Two ways a guess gives way to something known, each used where its condition holds. They answer different moments of the same field, in no set order.
In any order:
Fill from stored values — Autofill · optional
Where the system already holds the actor's own value, from a profile, an earlier session, or the browser, fill it in place of the guess. Do this only where confidence is high, since filled fields tend to be submitted unread.
Complete as the actor types — Autocomplete · optional
Where the valid answers are many and the actor is still typing, narrow the candidates to match. The candidates are looked up, not produced. Skip this where the answers are few enough to show in full.
Joining the work
Applies when
The guessed, stored, and looked-up values are in place, and what is left is work the actor would do themselves. The system can take part of it by producing content, which rests on inferred intent.
Read the intent from context — Embedded intelligence · optional
Infer what the actor is doing from what the system already holds, such as the selection, the last action, or the actor's role, so the actor never has to restate it to get help. The next step turns the inference into offers. Skip this where the actor states the intent, as in a conversation.
Make everything it produces an offer — Suggestion
Settle once the terms every higher level inherits: the system proposes and the actor decides. Accepting takes one gesture, declining is cheaper than weighing the offer, and each offer shows its confidence. An offer shown before the actor has formed their own idea narrows what they consider next, and checking how far they rely on it is handled elsewhere.
Continue what the actor is writing — AI completion · optional
Where the actor is composing, propose the continuation at the cursor, lighter than their own text and gone once they type past it. Stay silent when confidence is low. A block generated to be reviewed as a whole belongs with delegated results, not here. Skip this where nothing is being composed.
Proposing the next move
Applies when
So far the offers stay inside the step the actor is taking. The system can also see the points where one piece of work ends and another usually follows, and offer the next move.
Propose the next step — Next-best action
At a transition point, such as a closed record or a finished output, offer a whole action with the reason for it. These offers arrive uninvited, so spend the channel as carefully as the offers in it. If the system performs the action instead of proposing it, that is delegation, and this sequence ends there.
Guard the accept by what it does — Action consequences
Grade each proposed action by its consequences and size the guard to the grade: an easily reversed action keeps its one-gesture accept, and a consequential one needs a confirmation or a look at what it will touch. The grade also caps how far the level can rise later.
Adjusting the level of help
Applies when
Every level above the first section is a bet on typical use, and the bets age. Acceptances, dismissals, and corrections build up on each surface, and without a record, whether a level still fits is judged on impressions.
Keep a record of how offers are used
Record what was offered, what was accepted, what was dismissed and how quickly, and what was corrected after accepting, per surface and per level. This record measures how much help fits; the record of reliance measures how closely offers are checked.
Move each surface up or down — Assisted task completion
Raise a surface's level where its offers are routinely accepted and the grade allows it, and lower it where they are routinely dismissed. Decide the lowering too, rather than letting a level fall out of use. A higher level must explain more of itself, and routine acceptance never raises a consequential one past its grade.
Hand the recurring procedure over — Workflow
When the record shows the same proposals accepted whole in the same situations, turn them into a procedure both the actor and the system can read, with steps, conditions, and transitions, and hand it over to run. The actor now inspects and repairs the procedure instead of accepting offers, which leaves this sequence.
Leaves: Each surface sits at the level its record supports, and what recurred whole now runs as a procedure. The cost stays: every level rests on typical use, and it fits worst the actor who is least typical.
Connections
Calls Calibrating reliance
At Joining the work, step 2, via Cognitive forcing functions
an offer shown before the actor has formed their own view narrows it, and that sequence supplies the friction that answers it, at each offer surface's own grade
Calls Calibrating reliance
At Adjusting the level of help, step 2, via Transparent reasoning
a surface that moves up owes a fuller account of its offers, which that sequence's evidence steps provide
Hands off to Attention and tempo
At Proposing the next move, step 1, via Notification
proposals arrive uninvited at moments the system picks, and what may interrupt whom, and when, is that sequence's decision
Hands off to Delegating work, handing the work over, step 1
At Adjusting the level of help, step 3, via Workflow
the procedure is handed over as a standing request, one of the two ways into that sequence