Pinned Tweet
I made my software factory better
Now it has:
- GitHub integration for issues + PRs
- a deterministic pre-review gate
- built-in parallelism for single-agent orchestration
- Fable for intake and Haiku for reviews
… and more
It’s a true “v2” upgrade from my previous factory
youtube.com/watch?v=RiaTyJ2P…
what’s the new moat?
I think it’s reputation
(Particularly, a good one)
Ive been asking others about this, and a few have told me that it’s “trust”
Trust could be the new moat
That tracks for me.
Sure, you can vibe code DocuSign...
(This is real, by the way. Its called SignuDoc lmao)
But there’s no way I’ll trust it. Zero chance.
I doubt the guy who vibed it up even trusts it enough to use internally
Trust and reputation can’t be vibed. They dont come quickly or easily
They can only be earned through hard and consistent work
A word of warning however: these qualities have a disturbing asymmetry
They are far easier to lose than to acquire. And in the most extreme cases they can be decimated overnight
So protect them carefully
27
Voice prompting
Now minus the voice
‼️ OpenAI and Apple have both filed patent applications for wearables that read words you only mouth, without making a sound.
OpenAI's earbud senses tiny movements of your ear canal as your jaw moves. Apple's shines laser light on your cheek and tracks how the muscles shift.
Both filings describe two people wearing the devices and holding a completely silent conversation.
1
29
"understanding long term implications of technical choices"
^ hard skill that AI has not taken from SWEs
everyone coding with agents should internalize this -
the RLVR in current frontier LLMs' post-training is done at task level. in other words, they won't give a damn about long term maintainability
this is why we see so many useless comments and test cases being written by default. these things make future tasks less efficient, but they don't affect the current task's success, so they survived the training and became part of the models' behaviors
until post-training can reward long term maintainability (which i suspect is very hard because inherently it's a slow and expensive process to measure that), it's on us humans to correct it in the harness layer through prompts, skills, lint rules etc
this also implies that understanding long term implications of technical choices is a valuable human skill that's hard for an average person to make up for with LLM
it's the distinction of whether you can create a durable growing codebase, or have one that gradually degrades into insanity
if you have that skill, you are golden (for now, until post-training catches up)
1
21
“what I saw during the course of
progress so amazed me that I believe it is one dawn across science of something
far greater”
JUST IN: Riemann Hypothesis has been solved.
alpo.ge/remark.pdf
1
2
69
Alex ◯ ZazenCodes retweeted
Agreed. Use it
vercels deepsec harness is v. good and i highly recommend y'all run it over your services and code. thanks for making a great tool @cramforce
vercel.com/oss/deepsec
here's how to grow a newsletter exponentially
currently I'm at 2.5k subscribers
1,000 of those joined in the LAST MONTH
and I've been doing this for THREE YEARS
Same newsletter. Same format. Once a week, when I upload a new youtube video.
so, WHY NOW?
I'd love to tell you that it's "a thousand little things" or "getting better every day"
but the quality of my work has little to do with the success, I think
if I had to pick one reason why this is happening, it's this:
CONTENT STRATEGY
the way I see it, you need both
STRATEGY + QUALITY
if I could go back in time, I'd just tell myself to lock in on these. get obsessed with them.
But actually, I already knew this. and I've already been obsessed (yes, for years)
So really then, I would just tell myself to keep going.
to STAY obsessed.
5
5
221
I'm talking next week at @AITinkerers Toronto 🇨🇦
I'll hook up Jev to the SEC firehose and extract some data
why Jev?
2,300 SEC contracts are filed a month
that's 26M tokens (11K tokens each)
if you want to ask AI about each one:
GPT-6.2 Astra:
$225
6 hours
Jev
$1/month
15 minutes
yeah.
now, these numbers are just estimates. and the time assumes concurrent API requests.
But I've played with this data before.
And, in fact, some of the filings are much larger than 11k tokens.
I used a Meta SEC filing as an example dataset in my Jev course because it qualified as big context (outside of Jev's usual limits of ~64k tokens, thus requiring an extra header to be set)
((long-context mode in Jev can reach ~262k tokens))
needless to say, I'm excited to see what else Jev can do here and share it with a real live audience, in one of the greatest cities in the world
2
67
Alex ◯ ZazenCodes retweeted
Let me give my quick analysis of what I think is going on with all these outages / hacks / whatever they are.
I think it's a combination of a few things, all happening at once:
1️⃣ AI models and openly available offensive harnesses are being used to find new vulnerabilities, and some advanced attackers are getting far more dangerous as a result. This is happening at the capabilities level, but even more importantly the scale level. Meaning they can manage hitting way more targets now, and way more effectively
2️⃣ AI is also being used for not just vulnerability and exploit development, but recon, which can be quite noisy. It can also be potentially disruptive if it hits stuff that's fragile, especially if there's a lot of that traffic
🗒️ Quick Note: What a lot of people don't understand about "hacking" is that it's largely a whole lot of overhead. Mostly overhead, actually. Doing extensive research. Launching scans. Finding relationships between entities. Coming up with a target list. Enumerating tens of thousands of hosts and endpoints. Collecting massive amounts of data from those targets. Doing prioritization. Figuring out what we know about the defense. Customizing attacks for a particular target. Coming up with a strategy for the entire campaign. Keeping in mind being stealthy. Creating the attacks. Launching the attacks. Monitoring and managing them. Going through results. Etc. Etc. Basically, very little time is spent on the novel parts in the vast majority of attacks. It's 99.9% overhead, and it's hard and tedious and time-consuming to do that well. So there were very few attackers doing it.
THIS IS THE PART THAT AI JUST MADE 5-100X EASIER.
Anyway...
3️⃣ #1 and #2, combined with the media stories about them, are creating panic in companies and governments, which is causing them to start getting aggressive with patching. Even for stuff they haven't patched before....because it was too dangerous to production to patch them. Because....things might not go well, and things might break, and other things might depend on them...etc.
4. The combination of all these are producing outages. But it's extremely unclear how much of what we're seeing is the attacks vs. the patching, etc.
So basically, the internet is kind of an open wound. Always has been. We were always in a giant cave full of slithering creepy-crawlies, it's just the lights were out. And now everyone has flashlights they can shine wherever. And shoot whatever once they see it.
And now the flashlights and gunshots are flying everywhere. Except gunshots aren't the right metaphor. It's more like Orbital Space Cannons hitting giant parts of the internet at once. And anyone can have one! Fire up your model and offensive harness of choice and now you too have your own Orbital Space Cannon. Yay! You can point it at your bank, or your ex who cheated on you, or your company that fired you, or the utilities company that you hate. Or a city or country's critical infrastructure. Whatever.
Bottom line is that:
1. New attackers are coming online with these tools
2. Existing attackers are leveling up with them
3. That's resulting in way more vulns and attacks
4. This is forcing massive (and dangerous) patching
5. We're seeing outages
Again, we don't know how much attribution to give each cause, but this to me seems like the general situation. It could be that this is mostly patching-related. Or this could be mostly attack-related.
One thing I think it's NOT, is a single, major, coordinated attack by some attacker like a nation state. I think the analysis above is far more likely.
Anyway, hope this helps, and if you see an angle I missed please let me know.
Learning this word feels a bit supererogatory tbh
there's about half a million words in the english language
most aren't worth knowing. but many are.
so I built a learning system for claude code and codex that analyzes your prompts and finds gaps in your vocab
it will recommend words for you to learn, and then teach them to you with spaced repetition (anki)
it works in two parts:
1/ finding the word
2/ teaching the word
finding the word is the hard part.
you should know: this works for me, on my system. this is not something you can just pick up and use.
but, it is open source. here's how I'm doing it:
1/ a claude code / codex hook that runs when you submit a prompt
2/ a background agent looks at big prompts and tries to extract a new word for you to learn
3/ the words are stored in a cloudflare database and shared between an app on your desktop and your phone (mac / ios)
I might ship this as an actual app, so please let me know if you like the idea!
feel free to fork it and try it out yourself. to get started I'd use claude code or codex:
"adapt and install this <github_url> for my system as a desktop app only for now. you'll need to walk me through it. remind me to leave a star on the project"
github.com/zazencodes/precis…
4
50
Alex ◯ ZazenCodes retweeted
This company instantly became 10x more legit
We're excited to announce Firecrawl's Toronto office 🍁
It's our first office outside of SF, and we're hiring!
firecrawl.dev/careers
this post from David really resonated
it reminded my of dax's post from last week:
> i get it man everything i'm doing you were already doing a year ago i know i'm slow i'm late to everything and you're so ahead of the game but damn how come you're not more successful
nitter.cf/thdxr/status/212242565…
they are both very different but share a similar idea:
you are giving your attention to the wrong stuff.
case and point: the new Gemini and GPT models this week. these will not move your business forward.
but David's advice here? yup that sure as heck will
whose building tools for this? send me what you got.
while everyones focused on funsies skills and shit
the most useful thing you can do to accelerate your company is to make LLM generated artifacts easier to understand
whether thats a review walkthrough (read: not a mermaid diagram or more markdown text), or better visuals to explain things, thats whats needed to go faster and have trust
2
77
here's how to turn your Hermes agent into a software factory:
1/ install my factory skill
github.com/zazencodes/zazenc…
2/ add your repo
/factory add OWNER/REPO
3/ write a github issue describing what you want and give it the "factory" label
4/ run the factory
/factory run OWNER/REPO
hermes reads the issue, writes a spec, builds the change, runs the checks and opens a pull request
you review and merge it
5/ use the scheduling prompt in the repo to check for work every 10 minutes
- empty checks use no AI tokens (pre-agent startup script checks for open issues first)
- add issues to the queue and let it work through them
youtu.be/jSfK4VjDMEc
1
115
it's that time of the week...
my codex plan is about to reset (see usage in bottom left corner)
so I spun up a literal army overnight.
to do what, you ask?
1/ these guys edited a bunch of short-form content for me (using my own custom editing software)
2/ they updated my CRM
3/ they also made my vibe coded calendly app better (i use this to book calls, so I care about it working well)
I'm also working on a little something special. you can see if if you look near the top of my tmux-agent-plane sidebar
it's called zstack
I'm going to open source my stack. the personal ai system behind my work as a solo entrepreneur building and marketing online
2
3
279
“The tools you use change the way you think”
@CompleteSkeptic didn’t disappoint at ai engineer code last weekend
My favorite takeaways
- the big labs are shipping improvements (under immense pressure) on LLM architectures that may be fundamentally flawed, ie, RLHF
- step outside your tech bubble and you’ll notice that it’s not really impacting the real world much
- benchmarks are trash
- jev and decision 1 changed how we all think. Expect the way we think to keep changing (periodically , like a step function) as ai continues to advance
- the energy of the founder matters. I’ll bet on @typesafeai just because of Diogo’s energy and vision
Full talk: youtu.be/Cyt3LU8TyNc?si=oFov…
2
312
The jagged frontier of ai is so funny
Our best models can code at superhuman levels
…Solve millennium problems in math
But they are utterly incapable of writing a half decent X reply
The screenshot below is an interaction I had today. It’s nothing special. But it’s undeniably human
There is no ai system that would have created such an exchange of dialogue
(well, not Ben’s opening post which any ai could have slopped out easily. And perhaps did)
I imagine a future where this is solved. How it looks is that our AIs will learn to parrot us
Its sad actually, because it means we won’t ever know if we’re talking to a human or an ai anymore.
But from that pain will be born a solution. We’ll have human verified platforms where ai is banned. It seems hard or impossible now , but we’ll do it
The future, as always, is eternally bright. If you choose to see it
And indeed, how we see the future is a choice. And how we get there depends on more choices still
So choose wisely
58
there's about half a million words in the english language
most aren't worth knowing. but many are.
so I built a learning system for claude code and codex that analyzes your prompts and finds gaps in your vocab
it will recommend words for you to learn, and then teach them to you with spaced repetition (anki)
it works in two parts:
1/ finding the word
2/ teaching the word
finding the word is the hard part.
you should know: this works for me, on my system. this is not something you can just pick up and use.
but, it is open source. here's how I'm doing it:
1/ a claude code / codex hook that runs when you submit a prompt
2/ a background agent looks at big prompts and tries to extract a new word for you to learn
3/ the words are stored in a cloudflare database and shared between an app on your desktop and your phone (mac / ios)
I might ship this as an actual app, so please let me know if you like the idea!
feel free to fork it and try it out yourself. to get started I'd use claude code or codex:
"adapt and install this <github_url> for my system as a desktop app only for now. you'll need to walk me through it. remind me to leave a star on the project"
github.com/zazencodes/precis…
























