A draft comes back with a line you’d never have written. The draft is still sitting there waiting for you, and the list of what Point has learned is short and written in plain sentences. Two things put that line there. Point read the thread wrong, or Point is carrying out a sentence you said months ago, exactly as you said it. They look the same on the screen and they take opposite repairs, so the first move is to find out which one you have.
The second part of this series left Ellen in November with eight rules. Every one of them was right on the morning she said it. This part is the morning one of them produces a line she wants back. What she does now is a lookup, and a lookup has its own technique. It has almost nothing to do with judging whether the rules are any good.
The thing that goes wrong here is quiet. A draft that’s 90 percent right and 10 percent off gets repaired by hand in eleven seconds, sent, and forgotten. Everything about that morning looks ordinary. The standing instruction that just fired stays out of sight, and the line keeps running on every message of that kind until something makes you look.
Name the wrongness first
Ellen’s draft is a reply to Priya, a client of seven years, about the extension deadline on a partnership return. It reads:
Hi Priya,
March 17 is the deadline for the partnership, so I need the K-1 detail from you by the 10th at the latest.
Everything else is in.
Ellen
Every fact in it is right. The date is right, the ask is right, and it’s signed the way she signs things. The trouble is the opening. It puts a deadline in Priya’s face first, where Ellen would have spent a line on Priya. It reads like a firm she doesn’t run.
Now stop, because this is the part people skip. Before you open anything, say what’s wrong in the smallest concrete words you can, and describe the draft rather than your reaction to it. This is cold is a reaction, and so is this doesn’t sound like me. The first sentence is a deadline describes the draft, and that’s the one that works.
The reason is mechanical. You’re about to search a list of sentences, and a list of sentences answers to a property of the message. A mood matches every line. Start from this is cold and all eight come back as maybes, and you’re guessing again. Start from the first sentence is a deadline and you have one thing to match against. Most of the list falls away while you’re still reading it.
There’s a related habit in the first part of this series, which is naming your answer before you read the draft. That one gives you an independent position to check the draft’s facts against. This one gives you a search term.
Eight lines, forty seconds
Open the list. It’s plain sentences, one per line, newest at the top.
Every line is a trigger and an effect. The trigger is the audience and the occasion, the part of the sentence that says to clients I have worked with over a year or when the thread is about a deadline. The effect is what the finished message looks like once the rule has been followed. Two questions sort the list, in this order, and the first one is usually enough:
Could this line have fired on this message? Could it, rather than should it. A rule scoped to Ray stays silent on a reply to Priya, and you see that in under a second, before you’ve thought about what the rule does.
If it fired, would it produce the thing you named? A rule about sign-offs changes the last line of a message. A rule about how you open changes the first one.
Ellen’s eight go like this. The two about sign-offs fall on the second question, because they reach the last line and the trouble is in the first. The one about Ray falls on the first question. One of the two she said in a hurry is scoped to a particular engagement that closed in September, so it falls on the first question too, though it has now drawn attention to itself in a way that matters later. That’s four gone in about six seconds.
Four left. The rule about how she opens on a first reply falls away once she notices this is the ninth message on the thread. Three.
And there they are, the two about anything with a date in it, one of them written in the week she was behind:
Put the date in the first line of anything with a deadline in it.
She meant keep the date where Priya sees it. What she wrote names a position, and here’s the thing about position rules: a position is a scarce resource. Something already sat in the first line. To put the date there, the draft moved the greeting out of the way, which is what the sentence asks for. The rule she wrote has been followed to the letter.
Look at what just happened to the clock. Six lines went on their triggers, which is to say on the audience and occasion clauses that part two spent its length arguing you should write into the sentence. Those clauses were sold as a correctness measure, keeping a rule about clients away from your letter to the bank. This is the second payoff, and on an ordinary Tuesday it’s the one you feel. A rule with its edges on can be ruled out at a glance. A line reading only be less formal stays a maybe in every search you ever run, because it might have done anything. Vague sentences are slow to read, at exactly the moment you need the list fast.
The lines you did not type
Some lines in the list you typed. Others started as something you said in passing while looking at a piece of work. Point offered the sentence back, and you accepted it with a click. The list marks which is which, and during a lookup that mark tells you where to spend an extra two seconds.
Read the ones you didn’t type word by word, off the screen rather than from memory. What’s stored is the sentence that was proposed and you approved. The sentence you said out loud was a little different, and the gap between the two is usually small and occasionally the whole problem. You said something about deadlines going near the top. The list says first line. In November you’d have sworn those were the same sentence.
The proposing is doing its job here. Turning a spoken aside into something Point can carry out on messages you’re away from means making it specific, and specific is where a fair reading can land a step from what you meant. So the sentence on the screen is the one running your correspondence, and it’s the one worth reading during a search.
The lines you typed have their own hazard. You already know what they say, so your eye slides over them. If a line you wrote survives both questions, read it aloud. Five seconds, and it’s the cheapest way to catch a word doing more work than you gave it.
When nothing in the list could have done it
Most of the time the search comes back empty, and that’s an answer. It’s the answer to Ellen’s question, arrived at properly instead of guessed.
When every line falls on one of the two questions, the rules are clear and the thread is where this came from. Something was read out of the thread and read wrong: a date lifted from a message where it was still a proposal, a figure carried over from last year’s engagement, a name attached to the wrong company. The list is silent because the list was never involved.
Two things follow, and the second is the one worth holding onto.
The repair goes in the draft. You fix the sentence and send it, and the list stays as it is. What happened came out of one thread and it stays there, so a standing instruction has nothing to catch.
The other is the instinct to write a rule anyway, and that’s the one to resist. After a draft has embarrassed you, the natural move is to make sure it never happens again, and rules are the only lever in view. A sentence in the list can’t make a date correct. What it does is sit there permanently and slow down every search you run from here on, which is how a list that started at eight arrives at forty. The full sorting of what belongs in the list and what doesn’t, including the case where a rule is right and simply never reached the message, is worked through in seeing what an AI learned about you, and changing it.
There’s a second record for this case, and it answers a different question. The list tells you what Point believes. The activity log behind every action tells you what Point did, to which message, when, and on what reasoning. Go there when the question is why did this happen at all rather than why is it written this way.
Reading the list as one thing
Everything above treats the list as something you search. You can also read it whole. That read takes about the same forty seconds, and it tells you things a single line can’t.
Here are Ellen’s eight, stood back from: two about sign-offs, one about how she opens, one about Ray, two about dates, two from the week she was behind. Three patterns show up in that, and each one is invisible a line at a time.
A cluster is a rule you have yet to say. Two lines about sign-offs and two about dates are two things she cares about. She stated each one partly the first time and patched it the second. The real rule sits under the cluster, and it’s almost always simpler and wider than the sentences covering for it.
A person with a line to themselves is usually a relationship. Ray gets his own rule because Ray sits outside the client list and gets treated like a client. That line is correct, and it stays correct for exactly as long as Ray is at that company. It’s a piece of context wearing a rule’s clothes, and its expiry date arrives quietly.
The order is a record of what’s been bothering you. Newest at the top means the top of the list is your recent irritations. The bottom is what’s been running longest and quietly. Both ends tell you something. A line that’s sat at the bottom since March, drawing complaints from nobody, is doing its job. A cluster at the top is what you’re currently fighting with, and it usually turns out to be one sentence written badly rather than three subjects.
Then there’s the count itself. Eight is a list you can search in forty seconds. Forty is a list you leave closed on the morning you actually need it. The lookup stops happening, and the diagnosis stops with it. Why the short list is the working state rather than the unfinished one is set out in how Point learns your email habits.
The part that has no list
One question sits underneath all of this and the list leaves it open. It’s worth answering plainly, because owners go looking for a screen to answer it.
The rules you taught Point sit in one panel, written as sentences. The rest of what Point has picked up about your writing comes out of your own sent mail: years of your ordinary choices about greetings and length and how you say a date. That half is held as inference rather than as sentences, because it was never a sentence. There’s no list of it to open, no line to read and no line to remove.
That sounds like a gap, and here’s why it’s something else. The inferred half is the default, and the default is your own correspondence. A rule joins the list precisely because it departs from what your mail already shows, which is why the list is short and why it’s the only part with anything to look at. You read the inferred half in the drafts. A shape that keeps coming out slightly wrong on the same kind of message, with the list silent about it, is the inference showing through, and that’s a case for writing a rule.
Two other things Point learns live on their own surfaces, so change them there rather than in the memory panel. What Point learns from the way you correct a summary has its own subject and its own screen. Who counts as important to you is a mark on a person rather than a sentence about your writing, and you set it in day four of the first two weeks.
Where Point puts all of this
Point runs on the Gmail or Microsoft 365 mailbox you use now, so your address, your history and your sent folder stay where they are. The sent folder matters here more than it usually does. It’s the material underneath everything the list can leave unsaid.
Everything Point has been told sits in a single memory panel. The sentences are in your own language, the most recent one first, and the lines you typed carry a mark saying so. Any of them can be reworded or taken out. Putting a new one in is one plain sentence and nothing else. That panel is the whole store, which is what makes a forty-second search a real answer rather than a partial one: when the list is silent, it’s silent about everything Point holds.
A saved rule travels with the work, wherever Point is acting for you. A sentence about how you sign off to clients turns up under a message on a thread where sign-offs came up once, months ago, on a different thread. That reach is what makes one line worth writing, and it’s what makes this search worth running. A rule you stated in September stays quiet on a thread in November.
Alongside it, answering the other question, sits a written record of what was actually done on your behalf, in ordinary words, with each action one click from being put back. Undo reaches what’s still yours to reach. A message that has already landed on somebody else’s server sits outside that, and no arrangement of settings brings it back, which undoing what Point did is exact about.
How far Point goes on its own is a separate setting. It’s held for each type of work on its own and runs from suggest-only through review to fully handled. Drafting starts at review, where the words are prepared and the sending waits for you, and it stays there for that kind of work until you move it. The list and the dial are independent: emptying the list leaves the dial where it is, and filling it leaves the dial where it is. That decision is setting how much your inbox does on its own, made one kind of work at a time. The rest of what’s already finished by the time you sit down is on the benefits page.
Forty seconds later
Ellen has her answer, and it’s the uncomfortable one. The thread was read correctly. The line is hers. She wrote it in a week she was behind, it’s scoped correctly to messages with a deadline in them, and it did exactly what it says on a reply to a client of seven years.
It’s a good outcome, and it feels like a bad one. A found line is worth more than a repaired draft. The draft was one message. The line is every message of that kind between now and whenever she next notices. She noticed on a Tuesday in November, over something small. The other way to find out is in March, from a client.
Knowing which line did it leaves the repair open. The sentence is one word too specific, and the move that occurs to her first is to add a second line saying to keep the greeting. That’s the move that looks safest and costs most. Two overlapping sentences will both fire on one message one day, and she’ll be left guessing which one she’s arguing with. Editing what Point remembers about you is about the thirty seconds after the search, and about why the list gets shorter on a good day rather than longer.