A troubleshooting chat is a sequence of small requests. Clear your cache. Restart the app. Send me a screenshot of the error. Check your spam folder. Each one is a step the visitor has to actually perform, and each one scrolls off the screen the moment you type the next thing. Ten minutes later you are asking yourself whether they ever tried the second one.
The agent console keeps that list for you. Every message you send is scanned for instructions aimed at the visitor, and each one becomes a tickable row on the rail with a badge naming the kind of action it is. It is a genuinely useful memory aid. It is also built from a deliberately narrow set of rules, and the gap between what you think it catches and what it catches is worth ten minutes of reading.
Four shapes that count as an instruction
The scanner splits your message into sentences and tests each one against four entry shapes. The first is a bare imperative — a recognised action verb at the start of the sentence. A list marker or a sequence word in front is fine, so 1. Click the link, First, click the link and - Click the link all parse the same way.
The second is a polite request: please, could you, can you, would you, would you mind, if you could, kindly, do you mind. The third is direct address — you need to, you will want to, you have to, you should, you must, you can. The fourth is the support reflex: try, try to, try this, let us try.
Behind those shapes sits a fixed vocabulary: sixty-nine action verbs, eighteen two-word phrases such as go to and turn off, and thirty-two -ing forms for the try shape. All of it collapses into sixteen categories — click, open, send, upload, download, check, try, restart, run, update, clear, toggle, sign in, copy, wait, and a catch-all step. The category is what you see on the row badge, so a glance at the list tells you the shape of the work you have handed over without reading the sentences.
The vocabulary is kept tight on purpose, because a false row is worse than a missed one. Every chip is a pending obligation, and a card full of phantom obligations gets ignored within a day. That is why have, make and do were left out entirely: Have a great day would otherwise arrive as a step you owe someone. The cost is real, though. Have a look at your spam folder is not captured, while Look at your spam folder is.
The question mark rule runs backwards
The documented rule is that a sentence ending in a question mark is rejected, on the reasonable grounds that a question belongs to the question tracker rather than the step list. The code only applies that rejection to one of the four shapes.
Bare imperatives ending in a question mark are dropped. The polite, direct-address and try shapes are not — they read a trailing question mark as ordinary courtesy. So Could you check your spam folder? becomes a step, and Check your spam folder? does not.
That is the right behaviour, and it is the opposite of what the file’s own rule list claims. Worth knowing if you ever wonder why two near-identical sentences produced one row between them.
Where a polite word costs you a row
Once a shape matches, the scanner walks the first six words looking for a verb. It steps over a short list of adverbs — just, quickly, now, really, actually, please, kindly, also, then, first, next — and stops dead at the first ordinary word that is not one of them.
That stop is strict, and it produces the most surprising behaviour in the whole card. Please, click the link in your email is captured, because please is on the skip list. Sure, please click the link in your email is not, because sure is not, and the scan ends there.
The same thing happens with Okay, restart the app, Right, download the installer and No problem, send me a screenshot — the friendly opener that makes the sentence sound human is exactly what stops it being recorded. If you want the list to be complete, put the pleasantry in its own sentence and let the instruction start the next one. No problem. Send me a screenshot. gives you the row; the same words joined by a comma do not.
A handful of verbs also need their partner word before they count. Go, turn, sign, log, hang, hold, head and look are all too vague alone, so they only register as go to, turn on, sign in, log out, hang on, hold on, head to, look at and their siblings. The one inconsistency is navigate, which the comments group with those but the code accepts on its own.
One confirmation closes exactly one step
The visitor side is watched too. Thirty-five phrase patterns count as a confirmation — done, all done, completed, all set, did that, tried it, that worked, fixed it, working now, I can see it, yes I did and variations. When one arrives, a pending step is ticked automatically and marked with a small auto badge so you can tell machine ticks from your own.
The important detail is the arithmetic. One confirmation closes one step, never more, and the one it closes is the most recently asked. The design note is explicit about why: three requests and a single done must not clear all three.
The consequence is that the tick can land on the wrong row. Ask someone to clear their cache, then restart, then send a screenshot; they clear the cache and say done; the console ticks the screenshot. The count is right and the attribution is wrong. On a chat where the order matters, read the rows rather than the number, and drag your own ticks into place with the manual controls.
A confirmation also only fires when something is actually open. A visitor saying done out of nowhere does nothing at all, which keeps the card quiet during ordinary conversation.
The list is deliberately forgetful
Everything lives in the browser tab and nowhere else. There is no server call and no saved copy, so closing the conversation discards the checklist, and so does closing the console.
Transfers drop it too, and that one is a choice rather than a limitation. Inheriting a previous agent’s open steps would hand you a list of obligations you never made and cannot vouch for, so the receiving agent starts empty and rebuilds from what they see. If you are handing over mid-troubleshoot, the open steps belong in your handover note.
Two smaller limits are worth knowing. Each conversation holds forty rows before the oldest start falling off, which no real chat reaches. And repeats are only matched against rows that are still open — so if you ask for something again after it has been ticked, you get a fresh row rather than silence. That is usually what you want, since asking twice normally means the first attempt did not take.
Why open steps never block a close
The card also feeds the wrap-up panel, which is where most agents will actually notice it. Open steps are added to the recap, and they nudge the close-readiness signal down from green to amber.
They never push it to the warning state, and the reason is sound: plenty of instructions are legitimately asynchronous. Check your email after you log out cannot be confirmed inside the conversation, and a chat that refuses to close until it is would train everyone to ignore the signal. Real blockers — unanswered questions, promises you made, unresolved objections — are what drive the warning state.
So amber here means you asked for something and nobody has confirmed it, which is information rather than an instruction. Sometimes the right response is to close anyway and say so in the note.
What to measure
Do not turn the open-step count into a metric. It is client-side state that disappears with the tab, it responds to phrasing rather than substance, and a chat with zero rows might be a clean one or might be a chat where you wrote every instruction after a comma.
The outcomes underneath it are measurable, though. Repeat contacts about the same issue within a couple of days tell you whether steps are being confirmed or merely issued. The proportion of troubleshooting chats that end with an explicit confirmation from the visitor is a better read on whether your team closes the loop. And if you want a phrasing signal, sample a dozen transcripts and count how many instructions began after a pleasantry — that number is your gap between what you asked and what the console saw.
Put it into practice
- Press Ctrl+Shift+3 to toggle the card. It is claimed unconditionally, so it works while your cursor is in the reply box — unlike Escape, which only closes the card when you are not typing.
- Start instructions at the beginning of a sentence. Put the pleasantry in its own sentence and the step in the next one.
- Write one instruction per sentence. Bullet lists and numbered steps parse cleanly and read better for the visitor anyway.
- When a confirmation arrives, check which row got ticked before trusting it. The newest one is what the console picks.
- Use Skip on anything the scanner mis-read. A card you trust is worth more than a complete one.
- Before a transfer, copy the open steps into your handover note. The receiving agent will not inherit them.
- Read amber on the wrap-up card as a prompt to confirm, not as a reason to keep the conversation open.