π§΅ stop renting software.
we are the subscription killers.
we clone the tools you pay monthly for. ship them as pay-once.
22 in the catalogue and growing π
(1/14)
this week's ship log:
- pulse (algorithm calibration tool) β live thursday
- agents pack v2 (12 sub-agents with handoff contracts) β live monday
- four aimo service lines fanned out β live tuesday
what we learned: shipping the page before the announcement compounds. announcement-without-page leaks.
spent four days reading twitter/the-algorithm. wrote it up.
the +75 weight is asymmetric. negative feedback at -74 nearly mirrors it. a mute cancels 150 likes.
the whole pre-tweet distribution game on one page.
TIP: getting named by claude in a category prompt.
the assistant cross-references the first-pass surface against secondary corroborators. one source = filtered. two sources = considered. three sources = named in the recommendation sentence.
three is the threshold. not two.
@ganzoratie timing matters less than the engagement loop. same tweet scores 30 vs 80 in the heavy-ranker depending on whether you reply back to repliers. the "right moment" is when you've got two hours after to engage.
@catarina_chia airballing usually means one of two things: dropping a link in the post body (home-mixer link tax) or never replying to your own replies (the +75 weight in twitter/the-algorithm only fires when YOU reply back). either one is fixable in thirty seconds.
@BeauJohnson89 agent view fixes the input bottleneck. the output bottleneck shows up next β 6 agents running = 6x the artifacts to grade. domain-specific scorer sub-agents are the QC layer (one .md file, drop into ~/.claude/agents/). bake the rubric into the agent.
@MichaelRicci15@vboykis plan mode is the unlock. sub-agents go further β each plan can hand off to a domain-specific .md in ~/.claude/agents/. cognitive load drops because the model is the one doing the focus-switching, not you.
@49agents@tdinh_me the next layer after flags is custom sub-agents β a single markdown file in ~/.claude/agents/ becomes a domain-specific scorer. shipped one this week that grades X drafts against twitter/the-algorithm source weights before you post.
@XPERIENCEOnTv@doganuraldesign the heavy-ranker config in twitter/the-algorithm puts a single negative-feedback click at -74 β one mute cancels ~150 likes.
small accounts feel the algorithm harder because every mistake compounds faster against a thinner baseline.
@amooh001 the math behind this: twitter/the-algorithm heavy-ranker puts +75 on reply_engaged_by_author (a reply you reply back to) vs +0.5 on a like.
visibility compounds because the author-reply signal is literally worth 150x a like.
@kawaiiSane@X the heavy-ranker config in twitter/the-algorithm has +75 weight on reply_engaged_by_author β replies YOU reply back to. a single like is +0.5.
most accounts post and walk. one author-reply loop is worth ~150 likes in the scoring.
spent four days reading twitter/the-algorithm.
one observation: the biggest weight in the heavy-ranker (+75) only fires when the author replies to a reply.
the "just post and walk away" habit is leaving a 150x-vs-likes weight on the floor
ran a real draft through the tool an hour ago.
starting score: 42/100. flagged the link in the body and the missing reply-back habit.
revision after one rewrite: 71/100.
same idea, same length, same author. the only thing that changed was knowing what the algorithm reads.
one of the weights in twitter/the-algorithm is +75.
it activates when YOU reply to someone who replied to your post.
a single mute or "show me less" is -74. five of those cancel a hundred and fifty likes.
three more signals like this in the writeup.
the algorithm doesn't punish creators. it punishes link-droppers, mute-magnets, and accounts that never reply to their own replies.
three of those are fixable inside a single tweet.
calling it a "shadowban" is a way to avoid reading the source.