Akshat

building a ranked queue of X posts worth replying to - for cold accounts, not analytics for big ones

[Akshat @aksenHQ]

Growing a small X account is a reply problem first, then a posting problem.

Most people scroll for 20 minutes and find nothing worth a reply.

I rank posts by:
- freshness (still in the first hour)
- room left in the thread
- niche fit

That queue is what I'm building - for cold accounts, not analytics for accounts that already work.

Reply QUEUE if you want the list.

[Akshat @aksenHQ]

@CKPillaiAI Makes sense that the context part shipped first, since an agent picking up tickets is only as good as the ticket
Did they say what needs to be true before the autonomous part goes live?

[Akshat @aksenHQ]

@AmericaGetBig That's the second harness paper today where the same model jumps roughly 4x
Did they report cost per correct answer too, or only accuracy?

[Akshat @aksenHQ]

@uriel_builds $35 MRR across 4 customers works out to about $9 each
Is the low price on purpose to get feedback, or are you planning to raise it once you hit 10?

[Akshat @aksenHQ]

@mojtabajaveed If Financial Cents already ships it, at least you know accountants want it
What do their users complain about in reviews? It can be a sign of the gap to fix.

[Akshat @aksenHQ]

@0xverdao How did you measure the 24k it actually needed?
That's the number I'm curious about, since it’s hard to tell what the model used and what it just skimmed.

[Akshat @aksenHQ]

@sivagurudurai The part of the statement that caught my attention was "My job is mostly yes or no"
Given that you had 100 items in one sitting to consider, how did you decide which ones to give an actual look at?

[Akshat @aksenHQ]

@farrux_hewson Logging every action is a good habit. Before going to 100, I'd look at which of the 28 actually brought signups
Has any single channel produced a paying user yet?

[Akshat @aksenHQ]

@consumerxai Saves tell you more about demand than views do, since people only save things they plan to come back to
What did the 6 without app requests in the comments have in common?

[Akshat @aksenHQ]

@PeterAndersonE Japanese drops the subject a lot, so the model ends up guessing the pronoun on every page
Did a running character sheet fix it, or did you have to keep previous pages in context?

[Akshat @aksenHQ]

@alexander_jais Did they say what they tried it on? "Not good enough" often means they gave it a vague task in a big codebase they didn't know
Asking that could work as a screening question on its own

[Akshat @aksenHQ]

@codyschneider Getting the order right is the tricky part of a waterfall, because the cheapest provider usually isn’t the most accurate for every segment
Does the agent choose the order per lead or is it fixed?

[Akshat @aksenHQ]

@omarsar0 Going from 24.4 to 48.3 on GPT-5-mini with only harness changes is the bit most teams could copy today, no training needed
Did the evolved harnesses mostly add steps or remove them?

[Akshat @aksenHQ]

damn! how do you plan to use the subscription?

[Akshat @aksenHQ]

@evgnomon yikes!

[Akshat @aksenHQ]

@etelelorinczz The case study works because she genuinely is the ICP
did the trial reels reach non-followers better?

[Akshat @aksenHQ]

@MichaelGannotti 148/157 is high enough that the 9 misses are what I'd look at
Were they all the same kind of task or scattered?

[Akshat @aksenHQ]

@VGTimes If hand-written code is banned, every dev becomes a reviewer, and on the tricky parts reviewing code you didn’t write is slower than writing it. Did the slip start before or after the switch?

[Akshat @aksenHQ]

@colorosglobal @ManusAI On a phone the hard part is all the apps with no API
Will Mind Pilot act inside third-party apps or hand back a result for the user to finish?

[Akshat @aksenHQ]

@sunsetsyntax Keeping the fakes consistent is smart, since the model can still reason that staging-1 talks to db-2
What happens when it invents a hostname that isn't in your map?

[Akshat @aksenHQ]

@nmaties09 git already covers tracked files, so this mostly helps with the untracked stuff and .env
How does it know an agent started? Process name or something else?