Nouns Rise from the Surface, Verbs Stay in the Conversation
A copyable chip carried two things at once: a noun, the thing you meant, and a verb, what you wanted done with it. A month of real use split that cargo. The product now says what you are looking at, so “this” needs no carrying. The verb stays where it always belonged — in your own chat, where your authority and the record live. Along the way the slender window beside the chat became the app itself.
The Companion described a compromise and called it one. Host platforms give an outside application exactly one channel into the conversation — text the user pastes — so the ·Sum family made its bindings copyable: a chip whose face was the literal text it copied, a slender window beside the chat to hold the chips, and a three-mark grammar so the pasted text read as language to the person and as an exact reference to the agent. It worked, and I said I would retire it the day something better arrived. This essay is the record of that retirement. It did not come from the hosts. It came from the product learning to report what its member was looking at, and from a weekend in which I could no longer tell where I was inside my own software.
“Where am I?”
By late August Mail·Sum had three surfaces beside the chat: the slender Companion, the full website, and a pop-out reader for anything the Companion was too narrow to show. Each had a reason. Together they were experientially invisible. After a weekend of living in the product across desktop and phone I wrote down what I was experiencing — too many windows, no sense of wayfinding or state, taps on the phone minting orphaned tabs — and noticed that the founder was asking “where am I?” inside the thing he had built. The ruling that followed was blunt: one installed window on both form factors, and the app is the only door.
The shape is ordinary, which is the point. One window holds a rail beside a content pane. The rail carries the product's rooms, its pairing state, and its account menu. On a desktop the rail is pinned; below the breakpoint the same component becomes a drawer summoned under a slim sticky header — summoned, not shrunk. Every entry is a plain link and every move is real navigation, so the browser's own Back button is the wayfinding model. Reading happens in the content pane. There is no pop-out, no separate reader, no second product window of any kind. The web-app manifest carries the product's own name and its whole scope, so a link handed over in chat lands inside the one window instead of minting a tab. And a chat-requested act that changes what you will want to look at next returns an anchored link into the app as part of its result — always included, never offered, never asked, because discretionary phrasing drifts and asking is more friction than the link exists to remove.
What survived the ruling is more interesting than what it retired. The Companion essay said the family holds two standing surfaces, and beneath them two standing roles: a conversation plane for intent, authorship, and reasoning, and an instrument plane for orientation, selection, inspection, and authorization. Both survive intact. What changed is that the instrument plane stopped being a sidecar and became the product — full-size, rail plus content, home on Now. “Slender is the identity” was superseded, and every escalation from the slender panel to a reader window now reads, in hindsight, as the panel confessing it was too small for the work. The window was never the doctrine. The roles were.
Nouns rise from the surface
The deeper change had shipped two days earlier, and it is the reason the chips could go. Every product-owned surface in Mail·Sum now reports what the member is looking at — the message reader, the Sum page, the person page, and each of the product's screens — to the same server that renders them. It reports references, never content: an id, a kind, a view name, a version. One slot per owner, latest report wins, because a bare “this” means the freshest visible thing, not a merge of stale rooms. The reference carries the surface it rose from and its age, and the assistant is required to voice both when a pointer is more than a couple of minutes old. The cautionary datum behind that rule was a focus report thirty-three hours stale that was technically valid and conversationally false. A reference without a voiced age is a lie by omission.
Once focus rises from everywhere, the noun half of the chip is obsolete. You do not copy a citation for the page you are reading; the page has already said you are reading it. This is deixis done by the application rather than by the host: an application-owned semantic object with a stable reference, a version, and a repair lifecycle, instead of a screenshot or a highlighted DOM range the host guesses at. Every screenshot on this site's homepage is now a demonstration of it. “Reply to this.” “Tighten this paragraph.” “Give me a read on this photo.” Three products, three instruments, and not one handle, subject line, file name, or address in any of the prompts. The surface supplied the noun.
Verbs stay in the conversation
The tempting next step was to let the verb half follow the noun: turn the paste chips into toggles on the surface that the assistant reads — a checkbox meaning “summarize these,” a button meaning “compose a fresh reading.” It founders on an asymmetry that I think is general. Focus and selection are descriptive state: “I am looking at this; these are the ones.” Descriptive state ages gracefully — a stale pointer gets its age voiced and life goes on. A toggle meaning do this is imperative state, and a standing imperative in a database is a job queue wearing a toggle's clothes. It needs exactly-once consumption, ordering, and deduplication across conversations, because two chats that both read a pending “compose a reading” toggle will do the work twice. Mail·Sum met that monster in the same week under another name — the intent trap, where a page that merely reported focus could silently drain instructions meant for the paired window — and the cure was structural: reporters get a lane that cannot read, return, or mark an intent.
All of that machinery would buy me out of saying one short verb. And saying the verb is not a cost; it is the point. The request lives in the transcript, where my authority and the audit trail live, in my own words, in my own chat. So the doctrine is a pair of sentences. Nouns rise from the surface. Focus is the product's contribution — “this,” with provenance, version, and voiced staleness — and selection may contribute “these” wherever real use earns a plural (Mail·Sum's never did; a month of use never needed a plural once the singular was free). Verbs stay in the conversation. An action is a member-authored turn, and a surface control that means do this is refused on principle, not deferred for later.
Two residues, and one rule that stayed by evidence
Two kinds of paste survive, each for a reason the doctrine predicts. The pairing chip carries a secret, and carrying it is what proves which chat is mine; a toggle any conversation could read would pair any conversation. The paste is the authentication. Spelling pills — an @handle and its kin — are not requests at all; they are content for prose I am composing, and copy is their honest verb. Everything else that copied a citation retired product-wide, and the labels retired after them wherever the link already named the object.
One family of chips stayed for a reason the doctrine did not predict, and I keep it because the evidence was unambiguous. A few long requests — the pairing sentence, “update Now,” the two reading requests — carry composing rules with them. The same rules placed in a tool description were violated within minutes; pasted into the chat by the member, they were obeyed. A pasted request binds with the member's conversational authority, and the host treats it as the member's long sentence, pre-written. Their retirement trigger is written down and empirical: when spoken asks carry the rules as faithfully, twice in succession, the chips leave. Until then, they are the member's speech, not the product's.
Surface verbs execute
There is one more class, and it completes the grammar. Some verbs the product can finish without a model: reply, reply all, new message, write to a person. Those got real entrances on the surface, and they execute — into a member-owned artifact, immediately. Pressing Reply creates an actual draft in the signed-in product, one the member can write into directly or invite the assistant into midway (“help me write this”). Collaboration can begin from either side. The engineering rule that fell out is worth keeping with the grammar: one derivation, two callers. When the member and the assistant can perform the same act, give them one server-side implementation, so recipient rules, threading, and source linkage are shared architecture rather than two behaviors that tests must keep aligned. What is still refused is the deferred imperative — a verb waiting in server state for some assistant to pick up. Verbs either execute now, on the surface, or they are spoken, in the chat.
So Mail·Sum's affordance line, the sentence a designer on the family can hold in one hand, became: Labels name. Links navigate. Buttons execute. Focus points. Pills carry binding requests — and only those. Each clause divides work by authority and capability. Focus supplies nouns as descriptive state. Verbs the product can complete run into member-owned artifacts. Verbs that need the assistant's mind stay in the member's conversation, with binding chips serving as pre-written speech only where measured behavior justifies the carriage. Mem·Sum and Suminar still speak the earlier line — pills copy, links navigate, marks point — because their routed bindings still travel and their passage marks still feed edits. The grammar is one; the lines are product-specific applications of it.
The member's hands are the missing API
I can describe the future this approximates. I focus a window — a browser, an app — and my chatbot's own interface surfaces two or three contextually relevant actions for that object: “give me a fresh reading of this,” “stage a reply.” I tap one. That future does not overturn the doctrine; it preserves it. Tapping a surfaced prompt is still the member authoring a turn in their own conversation. Integration removes the typing, not the authority. And the protocol seam for it already exists: MCP prompts are exactly server-suggested action templates a host may show. Today's hosts render them as a static picker rather than keying them to focus. When a host closes that gap from its side, the product half is already built, because focus is precisely the signal a context-aware prompt surface would key on. Until then, the honest name for what we are doing is this: the member's hands are the missing API between surface and chat. The chips once made the member carry nouns and verbs both. Focus took the nouns. The verbs remain member-carried — briefly, by voice or by keystroke — which is where they should stay even after the carrying gets easier.
What it changes for the second bet
This site's second bet is that the rich surface develops beside the chat, not inside it as a widget, and that what has to develop between the two surfaces is a grammar, not a transport. This essay is the sharpest form of that claim I can currently make. WebMCP now lets a page hand the assistant its verbs, and Mail·Sum ships through that door inside ChatGPT's built-in browser beside the ordinary MCP connector — two entrances, one kernel. A tool list does not say which verbs belong on the page. The doctrine does. Verbs the product can complete into a member-owned artifact may execute on the surface; verbs that need the assistant's judgment stay in the chat; and any verb waiting in server state for an assistant to pick up is refused. The overnight probe recorded in The Grammar of the Second Surface showed why the last clause is not academic: given a page with no send tool, the host drove the browser to the button anyway. Ceremony for an irreversible act has to attach to the nature of the window, not to whose hand appears to click, because in an agent browser those hands share one channel.
It also settles what “the instrument plane” means once it is no longer a window shape. In Mail·Sum it is the one-window app. In Draft·Sum it is a pane inside Word, and the verb lands as a tracked change the writer accepts or rejects in Word's own Review interface. In Photo·Sum it is Lightroom itself, and the verb lands as a receipted step in the Develop panel. The plane is wherever the work already lives. What the family holds constant is not a place but a division: the surface says what is there, the conversation says what to do, and each side keeps the authority the other cannot forge (the case study).
The surface says what is there. The conversation says what to do. Neither carries the other's cargo.
Essay — building. The doctrine that retired the chip and the slender window, with the shipped surfaces that enact it. ← All essays