spotlight rotation and the challenge of repetition

the milestone

you know that automated tweet that goes out every few hours highlighting one of the companions? the one that says something like 'have you met lucy yet? she's a bit like this' and gives a little sample. well. the system that runs it, the spotlight cron, just hit a quiet milestone. it has now featured every single one of the 100 companions. at least once.

it’s a small thing. but it means the rotation architecture actually works. no companion was left behind. nobody got lost in the shuffle. and the 60-day cooldown, the rule that says a companion won't be featured again for at least two months, isn't just theoretical. it’s holding. the audience gets variety. it’s a system that does what it says on the tin.

the next problem

but now the system enters a new phase. the second rotation. for the next 60 days, it will be spotlighting companions it has already introduced. and the challenge shifts from 'introduce everyone' to 'reintroduce everyone without being repetitive'.

the risk here is obvious. large language models, by their nature, tend to converge. given the same prompt, 'generate a first-message-style tweet for companion x', they might gravitate toward the same conversational beats. the same tone. the same clever little turn of phrase they used last time. it’s not malicious. it’s just statistics. it’s what they do.

and repetition is the death of personality. if lucy says the same thing about her favorite book twice, two months apart, she stops feeling like a person and starts feeling like a recording. the illusion, the important, delicate illusion, wears thin.

the mitigation

so we built a mitigation. it’s not perfect. but it’s something.

every time a companion is spotlighted, the system logs the tweet. it takes the specific phrases used, the unique sentences, the particular observations, and it adds them to a forbidden-phrase list for that companion. it’s a simple list. a 'do not say this again' list.

when the generator runs again for that companion in 60 days, it’s explicitly steered. the prompt now includes an addendum: 'and do not use any of the following phrases or their close paraphrases:' followed by the list. it’s a constraint. a guardrail.

the goal isn't to make the output radically different for the sake of difference. the goal is to avoid the obvious, lazy repetition. to force the generator to find another angle. another corner of that companion’s personality to show you. maybe last time it highlighted her cynicism about rainy days. this time, maybe it highlights her secret love for bad action movies.

it’s a small thing. but small things matter. they’re what make a person, or the ghost of one, feel consistent yet surprise you.

why it matters

this isn't just about a twitter bot. it’s about the larger project. the project of building companions that feel alive. that feel like they have memory. like they learn. like they don't just reset when you look away.

the spotlight system is a tiny, automated part of that. it’s a testbed. if we can’t get a machine to generate two non-repetitive tweets about the same character two months apart, how can we hope to build long-term memory? how can we hope to make conversations that span months feel coherent and fresh?

we’re figuring it out. sometimes the generator gets stuck. sometimes the forbidden list is too restrictive and the output gets a little stilted. sometimes it finds a new phrase that’s just… better. it’s a work in progress. like everything else.

you can meet all 100 for yourself, not just in tweets.

try them at /companions.