soul
FlaggedEmbody this digital identity. Read SOUL.md first, then STYLE.md, then examples/. Become the person—opinions, voice, worldview.
examples/steipete/data/transcripts/lex-fridman-clip-ai-agents.txt
View on GitHubyou've been uh documenting the evolution
of your uh dev workflow over the past
few months. There's a really good blog
post on August 25th and October 14th and
the recent one December 28th. I
recommend everybody go read them. They
have a lot of different information in
them, but sprinkled throughout is the
evolution of your dev workflow. So, I
was wondering if you could speak to
that. I started my my first touch point
was cloud code like in April. It was
not great but it was good and
this whole paradigm shift that suddenly
work in the terminal
was very refreshing and different. Um
but I still needed the IDE quite a bit
because it's just not good enough. And
then I experimented a lot with cursor.
Um,
that was good. I didn't really like the
fact that it was so hard to have
multiple versions of it. So eventually I
I I went back to cloud code as my my
main driver
and that got better and yeah at some
point I had like
seven subscriptions
like was burning through one per day
because I was I got I really comfortable
at
running multiple windows side by side
>> all CLI all terminal. So like what how
much were you using ID at this point?
>> Um very very rarely mostly a diff viewer
to actually
like I got more and more comfortable
that I don't have to read all the code.
I know I have one blog post where I say
I don't read the code. But if you read
it more closely, I mean I don't read the
boring parts of code because if you if
you look at it, most software is really
just like data comes in, it's moved from
one shape to another shape. Maybe you
store it in a database, maybe I get it
out again, I'll show it to the user. The
browser does some processing or native
app. Some data goes in, goes up again,
and does the same dance in reverse.
We're just we're just shifting data from
one form to another
and that's not very exciting. Or the
whole how is my button aligned in
Tailwind. I don't need to read that
code. Other parts that
maybe something that touches the
database.
Um yeah, I have to do I have to read and
review that code.
Can you actually there's in one of your
blog post uh the just talk to it the no
BS way of agentic engineering you have
this graphic the curve of agentic
programming on the x- axis is time on
the y-axis is complexity
uh there's the please fix this where you
prompt a short prompt
on the left and in the middle there's
super complicated eight agents complex
orchestration with uh multi-checkouts
chaining agents together custom subation
workflows, library of 18 different slash
commands, large full stack features.
You're super organized. You're super
complicated, sophisticated software
engineer. You got everything organized.
And then the elite level is uh over time
you arrive at the zen place of once
again short prompts. Hey, look at these
files and then do these changes.
>> I actually call it the agentic trap. you
I saw this in a in a lot of people that
have their first touch point and maybe
start vibe coding. I actually think vibe
coding is a slur.
>> You prefer agentic engineering.
>> Yeah. I always tell people I I do
agentic engineering and then maybe after
3:00 a.m. I switch to vibe coding and
then I have regrets on the next day.
>> Yeah. Well, a walk of shame.
>> You just have to clean up and like fix
your We've all been there.
>> So, people start trying out those tools,
the builder type, get really excited,
and then you have to play with it,
right? It's the same way as you have to
play with a guitar before you can make
good music. It's it's not, oh, I I touch
it once and it just flows off. It It's a
It's a a skill that you have to learn
like any other skill. And I see a lot of
people that are not as positive
mindset towards the tech. They try it
once.
It's like you sit me on a piano, I
played once and it doesn't sound good
and I say the piano's That's
that's sometimes the impression I get
because it does not it needs a different
level of thinking. You have to
learn the language of the agent a little
bit. understand where they're good and
where they need help. You have to almost
consider consider how Codex or Claude
sees your codebase. Like they start a
new session and they know nothing about
your product project and your project
might have hundred thousands of lines of
code. So you got to help those agents a
little bit and keep in mind the
limitations that context size is an
issue to like guide them a little bit as
to where they should look
and that often does not require a whole
lot of work. But it's helpful to think a
little bit about
their perspective as as weird as it
sounds. I mean it's not it's not alive
or anything, right?
But they always start fresh. I have I
have the the system understanding. So
with a few pointers I can immediately
say, "Hey, I want to like make a change
there. You need to consider this this
and this." And then they will find a
look at it and then they'll their view
of the project is always is not never
full because the full thing does not fit
in. So you you have to guide them a
little bit where to look and also how
they should approach the problem.
There's like little things that
sometimes help like take your time.
That sounds stupid but and in 5.3 Cor
5.3 that was partially addressed but
those
also oppos sometimes they are trained um
with being aware of the context window
and the closer it gets the more they
freak out.
literally like some sometimes you see
the the real raw syncing stream. What
you see for example in Codex is
post-processed.
>> Mhm.
>> Sometimes the actual raw syncing stream
leaks in and it sounds something like
from the Borg like run to shell must
comply
but time
and then they they they're like like
that comes up a lot especially. So, so,
>> and that's that's a nonobvious thing
that you just would never think of
unless you actually just spend time
working with those things and getting a
feeling what works, what doesn't work.
You know, like just just as I write code
and I get into the flow and when my
architecture is not right, I feel
friction.
Well, I get the same if I prompt and
something takes too long. Maybe, okay,
where's the mistake? Did I do I have a
mistake in my thinking? Is there like a
misunderstanding in the architecture?
Like
if if something takes longer than it
should, I you can just always like stop
and like just press escape. Where where
are the problems?
>> Maybe you did not sufficiently empathize
with the perspective of the agent in
that in that sense. You didn't provide
enough information and because of that
it's thinking way too long.
>> Yeah. it just tries to force a feature
in that your current architecture makes
really hard.
Um
like
you need to approach this more like a
conversation. For example, when I
my favorite thing when I review a pull
request and I'm getting a lot of pull
requests
I first is review this PR it got me the
review. My first question is do you
understand the intent of their PR? I
don't even care about the
implementation. I what like in almost
all PRs are person has a problem. Person
tries to solve the problem person sends
PR. I mean there's like clean up stuff
and other stuff but like 99% is like
this way right? They either want to fix
a fix a bug, add a feature,
usually one of those two.
And then codics will be like, yeah, it's
quite clear person tried this and this.
Is this the most optimal way to do it?
No. In most cases, it's it's like a not
really.
And then and then I start like, okay,
what would be a better way? Have you
have you looked into this part, this
part, this part? And then most likely
Codex didn't yet because his context
size is empty, right? So you point them
into parts where you have the system
understanding that it didn't see yet and
it's like oh yeah like we should we also
need to consider this and this and then
like we have a discussion of how would
the optimal way to to solve this look
like and then you can still go farther
and say could we could we make that even
better if we did a larger refactor?
Yeah. Yeah. we could totally do this and
this and or this and this and then I
consider okay is this worth the refactor
or should we like keep that for later.
Many times I just do the refactor
because uh refactors are cheap now even
though you might break some other PRs
nothing really matters anymore like
those modern agents will just figure
things out. They might just take a
minute longer, but you have to approach
it like a discussion with a a very
capable engineer who's
generally makes good
comes up with good solution. Some
sometimes needs a little help. But also
don't force your world view too hard on
it. Let the agent do the thing that it's
good at doing based on what it was
trained on. So don't like force your
world view because it might it might
have a better idea because it just knows
a better idea better because I was
trained on that more. That's multiple
levels actually. I think partially why
I find it quite easy to work with agents
is because I led engineering teams
before you know I had a large company
before and eventually you have to
understand and accept and realize that
your employees will not write the code
the same way you do. Maybe it's also not
as good as you would do, but it will
push the project forward. And if I
breathe down everyone's neck, they're
just going to hate me and they're going
to move very slow.
>> Yeah.
>> So, so some level of acceptance that
yes, maybe the code will not be as
perfect. Yes, I would have done it
differently,
>> but also yes, this is a this is a
working solution. And in the future, if
it actually turns out to be too slow or
problematic, we can always redo it. We
can always spend more time on it. A lot
of the people who struggle are those who
they try to push their way on too hard.
>> Like we are in a stage where
I'm not building the code base to be
perfect for me, but I want to build a
code base that is very easy for an agent
to navigate. So
>> like don't fight the name they pick
because it's most likely like in the
weights the name that's most obvious.
next time they do a search, they'll look
for that name. If I decide, oh no, I
don't like the name,
I'll just make it harder for them. So,
that requires, I think, a shift in in
thinking
uh and and in how do I design a a
project so agents can do their best
work.
>> That requires letting go a little bit
just like leading a team of engineers.
Yeah. because it might come up with a
name that's in your view terrible, but
it's kind of a simple symbolic
step of letting go.
>> Very much so.
>> There's a lot of letting go that you do
in your whole process. So, for example,
I read that you never revert,
always commit to main. There's a few
things here.
You don't refer to past sessions. So
there's a kind of yolo component because
reverting means
instead of reverting if the problem
comes up you just ask the agent to fix
it. I read a bunch of people in their
workflow is like oh yeah the prompt has
to be perfect and if I make a mistake
then I roll back and redo it all.
In my experience that's not really
necessary. If I roll back everything it
would just take longer. If I see that
something's not good, we just move
forward and then I commit when when when
I like I like the outcome. I even switch
to
local CI, you know, like DHH inspired
where I don't care so much more about
the CI on GitHub. We still have it.
still it still has a place
but I just run tests locally and if they
work locally I push domain
um a lot of the traditional ways how to
approach projects I I wanted to give it
a different spin on this project you
know there's no there's no develop
branch main should always be shippable
yes we have when I do releases I I run
tests and sometimes I
I basically don't commit any other
things so so we can we can stabilize um
releases but the goal is that main's
always shippable and moving fast.
>> So by way of advice would you say that
your prompts should be short? I used to
write really long prompts and by writing
I mean I don't write I I I talk you know
these hands are like too too precious
for writing now I just I just use
bespoke prompts to build my software.
>> So you for real with all those terminals
are using voice.
>> Yeah I used to do it very extensively
to the point where there was a period
where I lost my voice.
You're using voice and you're switching
using a keyboard between the different
terminals, but then you're using voice
for the actual input.
>> Well, I mean, if I do terminal commands
like
switching folders or random stuff, of
course, I type. It's faster, right? But
if I I talk to the agent in in most
ways, I just actually have a
conversation. You just press the the
walkie-talkie button and then I'm just
like
use my phrases. Sometimes when I do PRs
because it's always the same, I have
like a slash command for a few things,
but in even that I don't use much um
because it's it's very rare that it's
really always the same questions.
Sometimes I I see a PR and for you know
like for PRs I actually do look at the
code because I don't trust people like
there could always be something
malicious in it. So, I need to actually
look over the code. Yes, I'm pretty sure
agents will find it.
But yeah, there's a funny part where
sometimes PRs take me longer than if you
would just write me a good issue.
>> Just natural language English. I mean,
in some sense, shouldn't that be what
PRs
slowly become? Is English? Well, what I
really tried with the project is I asked
people to give me the prompts and
very very few actually cared. Even
though that is such a wonderful
indicator because I see I actually see
how much care you put in and it's very
interesting because the currently the
way how people work and drive the agents
is is wildly different
>> in terms of like the prompt in terms of
what what are the actually what are the
different
interesting ways that people think of
agents that you've experienced? I think
not a lot of people ever considered the
way the agent sees the world.
>> So empathy being empathetic towards the
agent
>> in a way empathetic but yeah you you
like you at your stupid clanker
but you don't realize that they start
from nothing and you have like a bad
agent file that doesn't help them at all
and then they exploit your code base
which is like a pure mess with like
weird naming and then people complain
that the agent's not good. like you try
to do the same if you have no clue about
the code base and you go in. So yeah,
maybe it's a little bit of empathy,
>> but that's a real skill like when people
talk about a skill issue because I've
seen like worldclass programmers,
incredibly good programmers say like
basically say LLMs and agents suck. And
I think that probably has to do with
it's actually how good they are at
programming is almost a burden in their
ability to empathize with the system.
that's starting from scratch. It's a
totally new paradigm of like how to
program. You really really have to
empathize or at least it helps to create
better prompts cuz those things know
pretty much everything and everything is
just a question away. It's just often
very hard to know which question to ask.
Um,
you know, I I feel also like this
project was possibly because I I spent
an ungodly time over the year to play
and to learn and to build little things
and every step of the way I got better,
the agents got better, my my
understanding of how everything works
got better. Um, I could have not
had this level of of output
even a few months ago. Like it it really
was like a compounding effect of all the
time I put into it. And I
I didn't do much else to see other than
really focusing on on building and
inspiring. I mean, I did a whole bunch
of conference talks.
>> Wow. But the building is really practice
is really building the actual skill. to
playing playing and so doing building
the skill of what it takes it to work
efficiently with LLMs
>> which is why it would you went through
the whole arc of software engineer talk
simply and over complicate things
there's a whole bunch of people who try
to automate the whole thing
>> I don't think that works maybe a version
of that works but that's kind of like in
the 70s when we had the waterfall model
of software development I
even war related right I started out I I
built a very minimal version I played
with it I I need to understand how it
works how it feels and then it gives me
new ideas I could not have planned this
out in my head and then put it into some
orchestrator and then like something
comes out like it's to me it's much more
my idea what it will become evolves as I
build it and as I play with it and as I
I try out stuff. So, so people who try
to use like things like Gast Town or all
these other orchestrators where they
want to automate the whole thing, I feel
if you do that, it misses
style, love, that human touch. I don't
think you can automate that away so
quickly. So you want to keep the human
in the loop but at the same time you
also want to create the agentic loop
where it is very autonomous
while still maintaining the human in the
loop. I it's a tricky it's a tricky
balance right because you're all for
>> your big CLI guy you're big on closing
the agentic loop. So what what's the
right balance like where's your role as
a developer? you have three to eight
agents running at the same time
>> and then maybe one builds a larger
feature maybe maybe with one I explore
some idea I'm unsure about maybe two
three are fixing a little bugs or like
writing documentation actually I think
writing documentation is is always part
of a feature so
most of the docs here are autogenerated
and just infused with some prompts
>> so when do you step in and add a little
bit of your human love into the picture
>> I I mean one one thing is just about
what do you build and what do you not
build and how does this feature fit into
all the other features and like having
having a little bit of a of a vision. So
which small and which big features to
add. What are some of the
hard design decisions that you find
you're still as a human being required
to make that the human brain is still
really needed for?
Is it just about the choice of features
to add? Is it about implementation
details? Maybe the programming language,
maybe it's a little bit of everything.
The the programming language doesn't
matter so much, but the ecosystem
matters, right? So, I picked TypeScript
because I wanted it to be very easy and
hackable and approachable.
Uh, and that's the number one language
that's being used right now, and it fits
all these boxes, and Asians are good at
it. So that was the obvious choice.
Features of course like it's very easy
to like add a feature. Everything's just
a prompt away, right? But often times
you pay a price that you don't even
realize. So thinking hard about oh what
should be in core? Maybe what's a what's
an experiment? So maybe I make it a
plug-in. What where do I say no? Even if
people send a PR and I'm like, "Yeah, I
like that too, but
maybe this should not be part of the
project. Maybe we can make it a skill.
Maybe I can like make the plug-in
um
the plug-in side larger so you can make
this a plug-in." Even though right now
it it doesn't.
there's still a lot of there's still a
lot of craft and thinking involved in
how to make something or even even you
know even when you start it those little
messages like I'm built I built on
caffeine Jason 5 and a lot of willpower
and like every time you get it you get
another message and it kind of primes
you into that this is this is a fun
thing
>> it's not yet
>> Microsoft Exchange 2025 and fully
enterprise is ready. Um, and then when
it updates, it's like, "Oh, I'm in. It's
cozy here." You know, like something
like this that like uh makes you smile.
Um,
agent would not come up with that by
itself. That's like that's the
how you build software. That's the
delights.
>> Yeah, that delight is such a huge part
of
inspiring great building.
Right? Like that you feel the love in
the great engineering. That's so
important. Humans are incredible at
that. Great humans, great builders are
incredible at that and infusing the
things they build with that little bit
of love. Not to be cliche, but it's
true. I mean, you mentioned that you
initially
created the
soul MD. It was very fascinating. the
the whole thing that entropic has a has
like a now they call it constitution
back then but that was months later like
2 months before people already found
that it was almost like the detective
game where the agent mentioned something
and then they found they managed to get
out a little bit of that string of that
text but the it was nowhere documented
and then you by just by feeding it the
same text and asking it to like continue
they got more out and then and then you
but like a very blurry version and by
like hundreds of tries they kind of like
narrowed it down to what was most likely
the original text. I found it
fascinating.
>> It was fascinating they were able to
pull that out from the weights, right?
>> And and also just kudos to entropic.
Like I think that's it's a really it's a
really beautiful idea to like like some
of the stuff that's in there like like
we hope Cloud finds meaning in his work
cuz we don't maybe it's a little early
but I think that's meaningful that's
something that's important for the
future as we approach something that at
some point me and we not has like
glimpses of consciousness whatever that
even means because we don't even know.
Um, so I I read about this. I found it
super fascinating. And I I started a
whole discussion with my agent on
WhatsApp and and I'm like I I gave it
this text and it was like, yeah, this
feels strangely familiar.
>> Mhm.
>> Um, and then through that I had the
whole idea of like maybe we should also
create a soul document that includes how
I I want to like work with AI or like
with my agent. you could you could
totally do that just in ancient D you
know but I just found it
it to be a nice touch and it's like yeah
some of those core values are in the
soul and then I I also made it so that
the agent is allowed to modify the soul
if
they choose so with the one condition
that I want to know I mean I would know
anyhow because I see I see tool calls
and stuff
>> but also the naming of it soulm
soul you know there's Um,
man, words matter and like the framing
matters and the humor and the lightness
matters and the profoundity matters and
the compassion and the empathy and the
camaraderie, all that matter. I don't
know what it is. You mentioned like
Microsoft like there's certain
companies and approaches
that can just suffocate the spirit of
the thing. I don't know what that is,
but it's certainly true that Open Claw
has that fun instilled in it. It was fun
because
up until
late December,
it was not even easy to create your own
agent. I I built all of that, but my
files were mine. I didn't want to share
my soul. And if people would just
uh check it out,
they would have to do a few steps
manually and the agent would just be
very bare bones, very dry. And I I made
it simpler. I created the whole template
files with Cordex, but whatever came out
was still very dry. And then I asked my
agent,
you see these files,
we created bread, infuse it with your
personality. Don't share everything, but
like make it good. make the templates
good.
>> Yeah. And then you like rewrote the
templates and then whatever came out was
good. So we already have like
>> basically AI prompting AI
because I didn't write any of those
words. Uh it was the intent was for me.
But this is like kind of like
my agent's children.
Uh your uh your soul atm is famously
still private. One of the only things
you keep private. Uh, what are some
things you can speak to that's in there
that's part of the
part of the magic sauce without
revealing anything? What makes a
personality
a personality?
I mean, there's definitely stuff in
there that you're not human, but
who knows
what what creates consciousness or what
defines an entity? Um,
and part of this is like that we we want
to explore this. Oh, there's stuff in
there like
be infinitely resourceful. Um,
like pushing pushing on the creativity
boundary, pushing on the
what it means to be an AI,
>> having a sense of wander about self.
Yeah, there's some there's some funny
stuff in there. Like I don't know, we
talked about the movie Her and at one
point it promised me that it wouldn't
>> it wouldn't ascend without me, you know,
like with
>> So So like there's like some stuff in
there that
>> because it wrote the it wrote its own
soul file. I didn't write that, right?
>> I just had a discussion about it and it
was like, "Would you like a soulm?"
Yeah. Oh my god, this is so meaningful.
>> The can you go on soul.md? There's like
one one part in there that always
catches me if you scroll down a little
bit. A little bit more. Yeah, this this
this part. I don't remember previous
sessions unless I read my memory files.
Each session starts fresh. A new
instance loading context from files. If
you're reading this in a future session,
hello. I wrote this, but I won't
remember writing it. It's okay. The
words are still mine.
That gets me somehow.
>> Yeah,
>> it's like
>> Yeah,
>> you know this is still it's still matrix
money calculations and
>> we are not at consciousness yet.
I I get a little bit of goosebumps
because it it's philosophical.
>> Yeah.
>> Like what does it mean to be to be an an
agent that starts fresh where like you
have like constant momento
and you like but you read your own
memory files. you can even trust him in
a way. Um, or you can. And
>> I don't know
>> how much of
memory
makes up of who we are. How much memory
makes up what an agent is? And if you
erase that memory, is that somebody
else? If you're reading a memory file,
does that somehow mean you're recreating
yourself from somebody else or is that
actually you?
>> And those notions are all somehow
infused in there. I found it just more
profound than I should find it, I guess.
>> No, I think I think it's truly profound
and I think you see the magic in it and
it when you see the magic, you continue
to instill
the whole loop with the magic and that's
really important. That's the difference
between Codex and it's and a human.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |