Visitors stack questions. How much is shipping to Brazil? Do you ship on weekends? And what is the return window? arrives as one message, you answer the one you know best, and the other two scroll away. The visitor either re-asks or leaves thinking you dodged them.
The agent console keeps a per-visitor checklist of every question it can find, with an open count on the rail. It is a small feature with an unusually large effect on how a long chat feels from the other side. It also has one behaviour that surprises everyone who relies on it, and a couple of blind spots worth knowing before you do.
Two ways a sentence becomes a question
Every visitor message is stripped of markup, split into sentences, and each sentence is tested twice. A sentence that ends in a question mark qualifies immediately. A sentence that does not can still qualify if its first word is one of about thirty question openers — the what why when where who how which family, the modal verbs can could will would should shall may might must, and the auxiliaries is are am was were do does did have has had.
The opener path carries an extra condition: the sentence needs at least four words. That stops a bare How or a stray Do from becoming a row while still catching the very common unpunctuated ask — can you tell me the return window arrives without a question mark all day long, and it is caught.
Duplicates within a single message are collapsed, so a visitor who rephrases the same question twice in one breath gets one row rather than two.
The floor is lower than the notes claim
The documented minimum length is six characters. The actual minimum is four.
That difference is not academic. It is exactly the range where a bare Why? lands — four characters, ends in a question mark, so it bypasses the four-word rule and becomes a checklist row of its own. Under the documented rule it would have been dropped.
In practice this is a small win rather than a bug. A lone Why? in the middle of a support chat is often the most important thing the visitor says all conversation, and it is precisely the kind of one-word ask that gets talked over. Keeping it is the better behaviour. The lesson is about trusting the notes: the constant is the specification.
The contraction gap
The opener list is described as contraction-aware, with can’t you ship faster given as the example that still trips it. It does not.
There is no contracted form anywhere in the list, and the first word of a sentence is normalised in a way that keeps the apostrophe. So can’t is compared against the list as can’t, finds nothing, and the sentence falls through.
The gap only opens when the punctuation is also missing, because a question mark alone is enough to qualify. Can’t you ship faster? is captured on the punctuation rule. Can’t you ship faster is invisible — and so are won’t, isn’t, don’t, didn’t and haven’t in the same position. Those are the shapes an irritated visitor types, typed fast, without punctuation. Worth a manual glance on any chat that has gone tense.
What counts as one sentence
Sentence splitting needs whitespace after the terminal punctuation. That is almost always present, but visitors typing quickly on a phone routinely produce How much is shipping?Do you ship weekends? with no space, and that arrives as a single candidate — one row containing two questions, capped at two hundred characters with an ellipsis if it runs long.
One tick then covers two asks, which is the failure mode the card exists to prevent. It is rare enough not to worry about and common enough to recognise when a row looks oddly long.
There is a deeper fragility underneath. The splitter uses an invisible control character embedded directly in the source file as its delimiter. If that byte is ever stripped by an editor, a formatter or a copy-paste through a tool that normalises control characters, the split degrades to one candidate per character, every candidate fails the length floor, and the card goes silently empty rather than reporting an error. The team clearly knew: the newer step-tracking card explicitly abandoned the technique for a plain newline, with a comment saying it did so to avoid depending on a character that is fragile to retype. If your question pill ever sits at zero through an obviously question-heavy chat, that is the thing to suspect.
Any reply clears everything
This is the behaviour to internalise. A question is marked answered when an agent message arrives in that conversation after the question was asked. There is no matching between the reply and the question. There is no attempt to check that the reply is on topic, or even on subject.
So one message closes the entire open list. A visitor stacks three questions, you send One moment, let me check that for you, and all three flip to answered. The count you were relying on drops to zero while all three answers are still outstanding.
The card is honest about which ticks were automatic, and reopening one is a click. But the practical rule follows directly: the open count is only trustworthy up to your next message. Read the list before you send a holding reply, not after. If you need the list to survive, answer one question per message, or note the outstanding ones somewhere the console cannot helpfully tidy them away.
There is a reason the design is this blunt. Matching an answer to a question by content is a hard problem that fails in both directions, and a checklist that wrongly claims something is still open is more corrosive to trust than one that clears too eagerly. Knowing which way it errs is enough.
What it deliberately does not do
It never expands itself. A new question pulses the pill and nothing more, because the reply-box hint already surfaces the single most recent unanswered question. This card is the everything-else surface, and it stays out of the way.
State is per conversation and in memory. Forty rows per visitor before the oldest fall away, nothing written to a server, and a transfer or re-accept clears the list so the next agent is not handed a set of asks they never read. Manual controls — Done, Skip, Reopen — always outrank the automatic ticks.
It also only reads visitor messages. Questions you ask are not tracked here, which is the right split; that is a craft problem rather than a bookkeeping one.
What to measure
The open count is not a metric. It self-clears on your own reply, which makes it unusable as a management number and fine as a live prompt.
What you can measure is the shape of the problem it addresses. Sample transcripts and count the questions that received no substantive answer anywhere in the conversation — that is the real slip rate, and it is usually higher than anyone expects on chats with more than about eight visitor messages. Track re-asks as well: a visitor repeating a question is the clearest evidence available that the first one was lost. Both numbers move when agents change how they answer stacked questions, and neither depends on the card being right.
Put it into practice
- Press Ctrl+Shift+Q to open the card, then read the list before you send a holding reply. After it, the list will be clear.
- Answer stacked questions one message per question. The checklist stays useful and the visitor can follow which answer belongs to which ask.
- On a tense chat, scan for unpunctuated contractions yourself. Won’t you just refund it will not be on the list.
- Reopen anything a holding reply cleared. The auto tick is a suggestion, not a record.
- Use Skip on rhetorical questions so the count means something.
- Before transferring, restate the outstanding questions in your handover. The list does not travel.