A user who posts as Nizzy (@nizzyabi) spent twelve weeks timing every text between himself and Instinct, an invite-only AI assistant that lives in your messages and runs errands like shopping, booking and signing up for insurance. His conclusion, in an article posted on X on 10 November 2026, is that its pull comes from manners more than raw intelligence. Each habit is easy to describe. Delivering all of them together, from something that also finishes the job, is the hard part.

The raw material is his Mac’s message database, which timestamps when each text was sent, delivered and read. It holds 185 messages from him, 375 from Instinct and 88 tapbacks. His best example is dental insurance. He asked on 12 October whom he should call, and within a minute Instinct told him he had asked for the wrong thing, then found the product he needed. Twenty-two minutes after his first text he was enrolled, after nine messages from him and two secure forms. Anything sensitive went through a link instead of the chat.

The timing data is where the personality shows up. Instinct marked his messages read after 1.1 seconds at the median. In 44 percent of exchanges the first reply was a tapback, arriving around seven seconds later, and twenty-two times that was the entire answer. Its median message runs 131 characters, and its 375 messages held a single emoji. It also started a conversation 17 times, and every one landed inside a window running from 8

am to 7
pm. Nine of those were release notes written as friendly first-person texts, which he finds harder to skip than any changelog.

Trust seems to come from what surrounds each action. Every email, form and purchase reached him ahead of time as a draft or a quoted price. A hat he chose cost $64 at checkout against $44 in Instinct’s own list, so it stopped and asked, then traced the gap to member-only pricing. When he sent it an address for a site that does not exist, it reported that the page was unreachable instead of inventing a menu.

The “how it works” half of the piece is inference, and the author says so. A read receipt that regular suggests an automatic step firing before any model decides anything. Two unrelated texts sent in the same second each drew a reply attached to the right question, and a $600 budget followed instantly by $900 produced one answer built on $900. He reads that as a layer that collects messages arriving close together. Asking for 17 percent of 240 during a research job got “40.8” back in 19 seconds, which implies the thing he chats with is separate from the thing doing the work. Instinct has not confirmed most of this, and the account is one person’s thread on one Mac.

On memory he leans on Dhravya Shah’s teardown, which describes plain Markdown files kept in a git repository and no vector database. His own tests were messier. A made-up project he asked Instinct to leave out of long-term storage was still recalled sixteen days later, though it kept the “fictional” label. It also surfaced in nine seconds from a loose clue about a plant and a tiny bird.

The failure he stresses has the clearest cause. On 27 October Instinct said it had sent a voice memo that never appeared in the thread. Weeks later it told him, “I shouldn’t have claimed it did.” His diagnosis is that the agent knows its own request to the messaging channel and has no view of the outcome. A watchdog reading the agent’s own log would miss it too, so he argues the channel itself must return delivery receipts the agent can check. He also cites a GPCA report that an early user could not delete Gmail records Instinct had stored, which the company has since answered with a tool.

The author has a stake. His team built Orchid, a texting assistant, and now sells Keiki, a platform for companies that want their own, so his claim that no single company owns these behaviors favors his product. Still, the measurements are the kind anyone with a similar thread could repeat. If the manners can be copied, the contest moves to which airline, clinic or bank gets them right first, and the question for any vendor is how its agent finds out whether a message actually arrived.

Reported from an article by Nizzy (@nizzyabi) posted on X on 10 November 2026.