Chris Battarbee

letting ai do cool stuff safely | yc | ex-palantir

[Chris Battarbee @chrisjbattarbee]

Prompt I use when starting literally any task "[task here]. Ask questions until it makes sense"

This dramatically improves pretty much all my results

[Chris Battarbee @chrisjbattarbee]

@mattpocockuk Biggest benefit I found was to have it write the test suite prior to making the change

[Ece Kayan @ecekayan]

After having these conversations with myself (and @chrisjbattarbee) we decided to open source @withHumanAI under Apache 2.0.

Because lets face it. Anyone can copy anything with a single prompt. No license protects you from it.

[Chris Battarbee @chrisjbattarbee]

@sbalujaish Yeah the whole hand written weight changes is super interesting. I do wonder what the limit is, feels a bit like 2023 when we were fine tuning models but perhaps with a higher ceiling?

[Chris Battarbee @chrisjbattarbee]

> Trying to get opus 5.5 to write a coherent human readable sentence in 2026
> literally losing my mind

Is this what massive post training on coding causes? Do we need new models to solve this?

[Chris Battarbee @chrisjbattarbee]

@sbalujaish yeah perhaps I need to do a better TLDR 👀

[Chris Battarbee @chrisjbattarbee]

Turns out having an x presence is Xistential to running a dev tools startup these days - so here I am.

Like any engineer who’d rather read thousands of lines of Rust than write a single post, I went to the source code to see what shows up in your feed. Here’s what I found:

High level shit first:

1. Posts come from 4 different sources that get merged together into a list. Thunder (all the people they follow). Phoenix retrieval (AI prediction that the text you write matches their interest). SimClusters (People like that person liked it). Popular posts (top 500 popular posts across the platform)
2. All that gets merged into a mega list and filtered. Anything over 48 hours gets removed and the list is truncated to 2.8k entries
3. All the posts in that 2.8k list then get scored and ordered based on that score (more on that later)
4. Small account boosting moves some entries around 👀
5. Spam filters are applied
6. List is served and you can doomscroll to your hearts content

Three questions come out of this if you’re looking to show up in the feed
1. What can you do to show up in one of those 4 sources?
2. How does that ranking happen?
3. What is this small account boosting?

Getting into a source

Thunder: This is literally just a list of the most recent things that a person follows. This means literally just posting more gets you here more for your followers.

Takeaways: For me - I’m shit out of luck, I have no followers. But clearly having followers is OP.

Phoenix retrieval: a model reads your post's text and media and matches it to people whose recent activity fits.

Takeaways: Just write about stuff the people you want to reach are interested in

SimClusters: Bit more complicated. X is split into about 145k communities based on who follows who. Every account gets a list of communities it's into. When someone likes your post, their communities get added to your post's profile, and your post gets shown to other people in those communities.

So if your mum likes your post, it now gets shown to other people who love Gardeners' World and Knitting.

Takeaways: who likes your post is more important than just getting more likes. Make sure you’re getting people in the right communities for your post liking you, not just any people

Popular: Yeah we’re not making top 500 posts on x any time soon, let’s ignore this.

So if you’re reading this, your phoenix retrieval algo probably put it here

The ranking

Each likely action has a weight associated with it

Copy link: 20
Reply, quote, DM: 5
Follow you: 4
Share: 2
Repost: 1
Like: 0.66

Report: -234
Mute: -59
Not interested: -52
Block: -31

The model guesses how likely you are to do each thing, and the weights multiply those guesses.
So if the model thinks you have a 100% chance to like this post then the post has 0.66 added to its score.
If you have a 10% chance of copying the link then that adds 2 to the score.

How does it figure out the probability of you doing it?

it's a transformer (same kind of AI model as Grok) trained on a huge pile of "person was shown post, then did X" examples.

When it scores a post for you, it looks at:
- Your last ~1,000 actions: what you liked, replied to, shared, scrolled past, and how long you spent on each
- The post: what it's about (text and media), who wrote it, how old it is
- How the post is doing so far: likes, replies, reposts, quotes and views
- Whether you follow the author (or they follow you)
- Context like time of day and your language and location
Like scoring is logarithmic. Early likes count the most: going from 0 to 7 likes moves it as much as going from 1,000 to 7,000

So it's basically asking "people like you, who just did what you just did, how did they react to posts like this?"

Takeaways: the post's early likes and replies feed straight into everyone else's prediction, so a solid first hour snowballs.
Copy link is super OP (writing interesting shareable stuff seems good)

Small account boosting

If you have under 50k followers, a post less than 2 hours old with under 200 views in feeds can get lifted to around 15th place in someone's feed. This is good for people like me with nothing going on.

Takeaways: Theoretically if you could write good posts every 2 hours then that would be an optimal strat for someone with no followers

TL;DR for zero-follower me:

- One clear topic per post, so Phoenix knows who to show it to
- Get the right people liking it. Likes spread your post to the liker's communities (sorry mum, please don’t like this)
- Write stuff people want to send to a friend or reply to. Likes relatively suck
- Post when your people are online and be around for the first 2 hours, that's the boost window and is the window that determines if a post does well
A post is dead after 48 hours, it can’t be seen
Optimal posting schedule is something like once every 2 hours IF you can do it well
- Followers are OP. Every one of them gets your posts with no matching needed

[Chris Battarbee @chrisjbattarbee]

@dijedont are you using hermes rn? was looking for something to automate a bunch of recurring tasks, curious what your workflow is

[Chris Battarbee @chrisjbattarbee]

hermes writing code directly or hermes controlling claude code / codex?

[Chris Battarbee @chrisjbattarbee]

@ecekayan she's keener on gardeners world tbh

[Chris Battarbee @chrisjbattarbee]

@0xgolam All config, you set rules in your pipeline to define what you want to do with an approval request

[Chris Battarbee @chrisjbattarbee]

Sup everyone, we just launched withHuman today. Its an opensource approval layer for AI agents.

If you ever wanted to encode something like "refunds < $10 are done automatically but anything more I want to approve" then it's for you <3

Check it out https://t.co/wIpzOAHwxS

#opensource #AIAgents

[Chris Battarbee @chrisjbattarbee]

@ProductHunt We made opensource approval layer for AI agents.

If you ever wanted to encode something like "refunds &lt; $10 are done automatically but anything more I want to approve" then it's for you &lt;3

https://t.co/TZNUqzTzEZ

[Chris Battarbee @chrisjbattarbee]

@sebgoddijn Lets goooo! Congrats Seb, Alex &amp; the team

[Chris Battarbee @chrisjbattarbee]

@twtayaan Interesting approach for sure at the infrastructure level but lots more work that needs to be done at the application level!

[Chris Battarbee @chrisjbattarbee]

@tanayvk @browserbase @clerk https://t.co/ww2P0LeIoZ

[Chris Battarbee @chrisjbattarbee]

@TheGoDev library author here, lmk if you have any questions!

[Chris Battarbee @chrisjbattarbee]

@cesarvspr @GolangTrends Hey @cesarvspr , I'm the author of the library. It allows you to write tools that LLMs use so if you check out this https://t.co/DLH1mYoPsd demo, the LLM can use that tool to understand what the weather is like in different countries (by using the api call)