how it works

every follower on the wall was typed by an ai reading their public posts. no questionnaire, no oauth, no human in the loop. just a model, a prompt, and up to eighty of their most recent tweets.

this page explains what that actually means, where it cuts corners, and why the confidence chips exist.

the pipeline

scrape, prompt, score

1. the follower list comes from the public followers endpoint. each account's recent posts and replies are scraped once and stored, so re-running the prompt later costs nothing.
2. up to 60 original posts and 20 @-replies are fed to the model as evidence. replies are kept separate from posts because how someone talks to people is different evidence than how they broadcast.
3. the model scores each of the four axes independently, then the four-letter code is assembled from those picks. the code it writes is checked against its own axis picks, so a model that scores i,n,t,p and then writes "entp" cannot publish a code its own reasoning does not support.
4. every guess carries a quoted snippet from the person's actual posts. if the model cannot find something diagnostic, it says so instead of inventing a quote. quotes are validated after the fact by checking that the words actually appear in the corpus.

per-axis scoring

why not just ask for a type

asking a model for a single four-letter code lets a coin-flip axis silently decide the type and hides which one was weak. it also pulls toward entp, because "posts about tech" starts implying n once the whole type is one label.

scoring each axis separately, then assembling the code from those picks, cut the entp pull and made weak axes visible. that is what the confidence chips on each profile are: the model's own admission that it was guessing on at least one axis.

overall confidence is the weakest axis, not the average, so one coin flip is always shown.

medium bias

posting on x is not a personality

an uncalibrated model scores almost everyone as p and e, because posting on x is spontaneous, reactive, and public by nature. that is the platform, not the person.

the prompt corrects for this explicitly:

e/i is about where energy points, not volume. broadcasting to an audience is not extraversion. seeking back-and-forth and initiating with people is. long solitary analysis and treating the timeline as a notebook is introversion even at very high volume.
j/p is close to evenly split in the population. if the model scores almost everyone p, it is reading the medium. it is told to check itself, and given concrete j tells (ships on a cadence, finishes things, holds firm standards about process) and p tells (open loops, decisions revisited, projects abandoned midway).

before calibration, the j split was 21% across 1,592 people. after, it landed at 31%. reduced, not eliminated.

reproducibility

a single code is not a stable measurement

two runs of the identical prompt over the identical corpus, 40 profiles, agreed on the full four-letter type 65 to 70 percent of the time. per-axis agreement was high (90%+), but four axes compound: 0.9 to the fourth is about 0.66. roughly a third of the wall changes type on a rerun.

temperature 0 does not fix it. the instability is in the task, not the sampling. this is the strongest argument for the confidence chips and for the quiz: a human answer becomes the tiebreak on something genuinely contested.

the quiz

prove it wrong

there are two ways in, and the order decides when your result goes public:

quiz first. take the 24-statement quiz before claiming. your result is private: only your browser can see it, and the ai guess stays on the wall. when you redeem the claim link from your dms, that same result is promoted to public and replaces the ai guess on your card and profile page.
claim first. redeem the dm link, then take the quiz. your result publishes immediately, since the claim already proved the account is yours.

claiming is one click from a dm. no email, no password, no oauth. the server stores only a hash of the claim token, not the token itself, and the cookie it sets only identifies you to this site.

what it is not

not a diagnosis. myers-briggs is a framework, not a clinical instrument, and an ai reading tweets is not a therapist.
not permanent. types are guesses until someone claims them, and the ai guess can change on a re-run.
not affiliated with the myers-briggs foundation. it is a toy.