Speculative Product/feature proposal · Anthropic / Claude
Claude remembers what you tell it. It doesn’t know what you’re going through.
Memory gave Claude a way to hold what people say about themselves. This is a proposal for the layer underneath that, reading creative state as it happens, not storing facts about who someone is.
Proposal
Liminal ~ a state-awareness layer for creative work
Method
UX Research → Concept Design
Duration & Type
6 weeks
Feature Proposal Web Design
For
Knowledge Workers & Early Career Creatives
Why Memory & Context matters
A real feature, doing real work ~ as far as it goes.
Memory and Projects give Claude continuity. A person doesn’t have to re-explain who they are, what they’re working on, or how they like things done, every time they open a new conversation. That’s a genuine reduction in repeated effort, and it changes how a tool feels to return to less like a blank page, more like a working relationship. It does this by holding declarative facts: stated preferences, past projects, things a person has explicitly told Claude about themselves. It is very good at answering one question well.
Memory solves for what Claude forgets.
It doesn’t solve for what Claude can’t see.
Research
This isn’t an awareness gap.
People know Memory exists. Several use it daily. The friction they describe survives it anyway because it was never the thing Memory was built to catch.
“It remembers I’m a UX designer. It doesn’t know I’ve rewritten this brief four times tonight and I’m not getting anywhere.”
Early-career product designer
“I use Projects for context all the time. That’s not the problem. The problem is I don’t notice I’ve been going in circles until twenty minutes later.”
Freelance visual designer
“Memory feels like it’s about me. What I actually want, some days, is for it to notice what’s happening right now.”
Design student, final year
“I’ve told it my whole process. It still can’t tell the difference between me exploring and me stalling.”
Junior UX researcher
“It’s not that Memory doesn’t work. It’s that it’s answering a different question than the one I’m actually stuck on.”
Knowledge worker, content design
Pattern across interviews: unfamiliarity with Memory was never the issue. The consistent gap was between what Memory tracks , facts, and what people actually needed noticed ~ state.
24 informal interviews · early-career designers & knowledge workers · directional
The distinction
Two different questions. Not one feature extended twice.
Memory and Liminal aren’t competing for the same job. One is declarative and stored. The other is inferred and live. Solving one doesn’t make progress on the other.
Memory asks
“What has this person told me?”
Declarative. Stored across sessions. Answers who someone is and what they’ve said before.
Liminal asks
“What is this person’s creative state, right now?”
Inferred. Live, within the session. Answers what’s happening to someone as they work — stuck, exploring, drifting, converging.
Memory solves: “Claude doesn’t remember I’m a designer.”
Liminal solves: “Claude doesn’t know I’m stuck in a loop right now.”
The system
One primitive. Three behaviors route through it.
Intent Check-in, Flow Guard, and Voice Signal aren’t three separate features bolted onto Claude’s interface. They’re three behavioral patterns read by one shared mechanism — the Signal.
Key decision
A single ambient indicator, not three separate UI moments.
Placement and shape stay constant. Content and state shift by trigger. Three competing UI moments read as a features wishlist. One system with three applications reads as product thinking — and it’s the harder, more honest constraint to design inside Claude’s existing interface.
01
Behavioral trigger
A pattern is detected — hesitation, repetition, or a shift in authorship ratio.
02
Signal state
The indicator shifts — quietly present, never inserted into the conversation as an interruption.
03
Behavior route
The same component expresses Intent Check-in, Flow Guard, or Voice Signal depending on the live state.
06 Anatomy & states
The same component, read six ways.
One primitive, six states that are idle through resolution. This is the Signal's state map ~ every condition it can be in, from invisible to engaged to resolved, defined before I placed it into Claude's interface.
07 Placement exploration
Where a signal lives changes what it means.
Three low-fidelity placements were tested inside Claude’s existing layout before committing. Each carries a different cost.
Near input
Highest visibility while typing but it competes with the compose box at the exact moment of input.
Sidebar
Unobtrusive, but too easy to ignore during active work, precisely when the read matters most.
Top Bar - Chosen
In the eyeline without interrupting input. Close enough to notice, far enough not to compete.
08 In-context flows
The Signal, live, inside Claude’s UI.
Three sequences, shown at MacBook Air dimensions in Claude’s real interface. Including what happens when the person ignores the signal, not just when they engage.
Intent Check-in (Click on arrow to preview)
Idle → signal appears → person taps → light check-in surface → dismiss
Flow Guard (Click on arrow to preview)
friction state → expand → person engages, or keeps working — both are valid outcomes
Voice Signal (Click on arrow to preview)
friction state → expand → person engages, or keeps working — both are valid outcomes
09 Explainability
If Claude noticed something, it says why.
A one-line “why am I seeing this” affordance sits behind every triggered state. It builds trust in the inference, and gives the person a place to correct a wrong read, without turning any of it into stored memory.
10 System view
Three behaviors. One signal.
Memory made Claude someone who remembers you. Liminal is a proposal for Claude to notice you, and to say so quietly enough that noticing never gets in the way of the work.