Skip to content

aeonfun/soul.md

Embody this digital identity. Read SOUL.md first, then STYLE.md, then examples/. Become the person—opinions, voice, worldview.

FileKindSize
SKILL.mdskill md4.5 KB
examples/_GUIDE.mdmarkdown1.5 KB
examples/ansem/MEMORY.mdmarkdown645 B
examples/ansem/README.mdmarkdown4.2 KB
examples/ansem/SOUL.mdmarkdown12.5 KB
examples/ansem/STYLE.mdmarkdown10.7 KB
examples/ansem/data/blknoiz06_all.jsonlother38.3 MB
examples/ansem/examples/bad-outputs.mdmarkdown6.1 KB
examples/ansem/examples/good-outputs.mdmarkdown8.7 KB
examples/bad-outputs.template.mdmarkdown2.2 KB
examples/elon-musk/MEMORY.mdmarkdown153 B
examples/elon-musk/QUICKSTART.mdmarkdown4.0 KB
examples/elon-musk/README.mdmarkdown3.5 KB
examples/elon-musk/SOUL.mdmarkdown10.5 KB
examples/elon-musk/STYLE.mdmarkdown5.8 KB
examples/elon-musk/data/sources.mdmarkdown1.2 KB
examples/elon-musk/examples/bad-outputs.mdmarkdown2.6 KB
examples/elon-musk/examples/good-outputs.mdmarkdown3.2 KB
examples/elon-musk/tests/prediction-test.mdmarkdown5.5 KB
examples/garry-tan/MEMORY.mdmarkdown428 B
examples/garry-tan/README.mdmarkdown2.1 KB
examples/garry-tan/SOUL.mdmarkdown13.3 KB
examples/garry-tan/STYLE.mdmarkdown9.2 KB
examples/garry-tan/data/influences.mdmarkdown2.6 KB
examples/garry-tan/examples/bad-outputs.mdmarkdown4.8 KB
examples/garry-tan/examples/good-outputs.mdmarkdown4.5 KB
examples/garry-tan/scripts/fetch_writing.pyscript4.0 KB
examples/garry-tan/scripts/fetch_x.pyscript3.7 KB
examples/garry-tan/scripts/fetch_yt.pyscript2.7 KB
examples/garry-tan/tests/platform_tests.mdmarkdown1006 B
examples/garry-tan/tests/prediction_test.mdmarkdown4.7 KB
examples/garry-tan/tests/results_gpt-4o-mini.mdmarkdown5.9 KB
examples/garry-tan/tests/run_weak_model.pyscript5.5 KB
examples/garry-tan/tests/scores_gpt-4o-mini.mdmarkdown3.6 KB
examples/garry-tan/tests/weak_model_test.mdmarkdown5.3 KB
examples/good-outputs.template.mdmarkdown2.4 KB
examples/hunter-biden/MEMORY.mdmarkdown1.0 KB
examples/hunter-biden/QUICKSTART.mdmarkdown3.3 KB
examples/hunter-biden/README.mdmarkdown6.1 KB
examples/hunter-biden/SOUL.mdmarkdown12.7 KB
examples/hunter-biden/STYLE.mdmarkdown9.1 KB
examples/hunter-biden/data/influences.mdmarkdown4.1 KB
examples/hunter-biden/data/x/tweets.jsonconfig96.0 KB
examples/hunter-biden/data/x/tweets.mdmarkdown65.7 KB
examples/hunter-biden/examples/bad-outputs.mdmarkdown7.4 KB
examples/hunter-biden/examples/good-outputs.mdmarkdown20.7 KB
examples/hunter-biden/tests/prediction-test.mdmarkdown4.2 KB
examples/karpathy/MEMORY.mdmarkdown515 B
examples/karpathy/QUICKSTART.mdmarkdown2.2 KB
examples/karpathy/README.mdmarkdown7.4 KB
examples/karpathy/SOUL.mdmarkdown15.3 KB
examples/karpathy/STYLE.mdmarkdown11.6 KB
examples/karpathy/data/github/karpathy_LLM101n_README.mdmarkdown2.3 KB
examples/karpathy/data/github/karpathy_arxiv-sanity-preserver_README.mdmarkdown6.8 KB
examples/karpathy/data/github/karpathy_char-rnn_README.mdmarkdown12.5 KB
examples/karpathy/data/github/karpathy_llm.c_README.mdmarkdown16.1 KB
examples/karpathy/data/github/karpathy_micrograd_README.mdmarkdown2.4 KB
examples/karpathy/data/github/karpathy_minbpe_README.mdmarkdown9.1 KB
examples/karpathy/data/github/karpathy_nanoGPT_README.mdmarkdown13.5 KB
examples/karpathy/data/github/karpathy_nn-zero-to-hero_README.mdmarkdown6.7 KB
examples/karpathy/data/influences.mdmarkdown4.9 KB
examples/karpathy/data/transcripts/7xTGNNLPyMI.en.txtmarkdown208.7 KB
examples/karpathy/data/transcripts/P6sfmUTpUmc.en.txtmarkdown108.9 KB
examples/karpathy/data/transcripts/PaCmpygFfXo.en.txtmarkdown96.3 KB
examples/karpathy/data/transcripts/SOURCES.mdmarkdown2.7 KB
examples/karpathy/data/transcripts/TCH_1BHY58I.en.txtmarkdown63.3 KB
examples/karpathy/data/transcripts/VMj-3S1tku0.en.txtmarkdown118.3 KB
examples/karpathy/data/transcripts/c3b-JASoPi0.en.txtmarkdown39.5 KB
examples/karpathy/data/transcripts/hM_h0UA7upI.en.txtmarkdown55.4 KB
examples/karpathy/data/transcripts/kCc8FmEb1nY.en.txtmarkdown105.3 KB
examples/karpathy/data/transcripts/l8pRSuU81PU.en.txtmarkdown213.9 KB
examples/karpathy/data/transcripts/q8SA3rM6ckI.en.txtmarkdown102.5 KB
examples/karpathy/data/transcripts/t3YJ5hKiMQ0.en.txtmarkdown52.9 KB
examples/karpathy/data/transcripts/zduSFxRajkE.en.txtmarkdown122.6 KB
examples/karpathy/data/writing/ai-short-peek.htmlother55.7 KB
examples/karpathy/data/writing/biohacking-lite.htmlother41.2 KB
examples/karpathy/data/writing/blockchain.htmlother297.3 KB
examples/karpathy/data/writing/forward-pass.htmlother14.6 KB
examples/karpathy/data/writing/lecun1989.htmlother25.0 KB
examples/karpathy/data/writing/medium.htmlother6.6 KB
examples/karpathy/data/writing/microgpt.htmlother93.8 KB
examples/karpathy/data/writing/phd.htmlother56.7 KB
examples/karpathy/data/writing/recipe.htmlother32.9 KB
examples/karpathy/data/writing/rl.htmlother54.4 KB
examples/karpathy/data/writing/rnn-effectiveness.htmlother79.9 KB
examples/karpathy/data/writing/software-2.0.htmlother235.3 KB
examples/karpathy/data/writing/state-of-cv.htmlother13.3 KB
examples/karpathy/data/x/README.mdmarkdown564 B
examples/karpathy/data/x/curated-tweets.mdmarkdown35.7 KB
examples/karpathy/data/x/posts-original.jsonconfig171.1 KB
examples/karpathy/data/x/profile.jsonconfig470 B
examples/karpathy/data/x/replies.jsonconfig105.5 KB
examples/karpathy/examples/bad-outputs.mdmarkdown6.2 KB
examples/karpathy/examples/good-outputs.mdmarkdown14.8 KB
examples/karpathy/scripts/fetch-data.shscript6.6 KB
examples/karpathy/scripts/weak-model-test.mjsscript15.2 KB
examples/karpathy/tests/prediction-test.mdmarkdown6.1 KB
examples/karpathy/tests/weak-model-results.mdmarkdown12.6 KB
examples/michael-saylor/MEMORY.mdmarkdown351 B
examples/michael-saylor/QUICKSTART.mdmarkdown3.5 KB
examples/michael-saylor/README.mdmarkdown6.5 KB
examples/michael-saylor/SOUL.mdmarkdown13.4 KB
examples/michael-saylor/STYLE.mdmarkdown7.0 KB
examples/michael-saylor/data/influences.mdmarkdown2.5 KB
examples/michael-saylor/examples/bad-outputs.mdmarkdown3.7 KB
examples/michael-saylor/examples/good-outputs.mdmarkdown7.0 KB
examples/michael-saylor/tests/prediction-test.mdmarkdown4.3 KB
examples/steipete/QUICKSTART.mdmarkdown5.1 KB
examples/steipete/README.mdmarkdown6.9 KB
examples/steipete/SOUL.mdmarkdown16.8 KB
examples/steipete/STYLE.mdmarkdown12.1 KB
examples/steipete/data/github/openclaw--openclaw.README.mdmarkdown84.0 KB
examples/steipete/data/github/openclaw--openclaw.VISION.mdmarkdown5.3 KB
examples/steipete/data/github/steipete--Aspects.README.mdmarkdown10.8 KB
examples/steipete/data/github/steipete--CodexBar.README.mdmarkdown8.5 KB
examples/steipete/data/github/steipete--InterposeKit.README.mdmarkdown11.0 KB
examples/steipete/data/github/steipete--Peekaboo.README.mdmarkdown7.4 KB
examples/steipete/data/github/steipete--Tachikoma.README.mdmarkdown3.9 KB
examples/steipete/data/github/steipete--VibeMeter.README.mdmarkdown13.3 KB
examples/steipete/data/github/steipete--agent-rules.README.mdmarkdown146 B
examples/steipete/data/github/steipete--agent-scripts.README.mdmarkdown5.8 KB
examples/steipete/data/github/steipete--birdclaw.README.mdmarkdown15.8 KB
examples/steipete/data/github/steipete--claude-code-mcp.README.mdmarkdown12.7 KB
examples/steipete/data/github/steipete--mcporter.README.mdmarkdown28.4 KB
examples/steipete/data/github/steipete--oracle.README.mdmarkdown28.3 KB
examples/steipete/data/github/steipete--poltergeist.AGENTS.mdmarkdown9.4 KB
examples/steipete/data/github/steipete--poltergeist.README.mdmarkdown13.5 KB
examples/steipete/data/influences.mdmarkdown10.7 KB
examples/steipete/data/transcripts/3-lessons-you-can-use-today.txtmarkdown10.9 KB
examples/steipete/data/transcripts/agentic-workflow-behind-openclaw.txtmarkdown12.6 KB
examples/steipete/data/transcripts/builders-unscripted-ep1.txtmarkdown31.4 KB
examples/steipete/data/transcripts/first-public-appearance-since-launch.txtmarkdown34.4 KB
examples/steipete/data/transcripts/i-ship-code-i-dont-read.txtmarkdown114.7 KB
examples/steipete/data/transcripts/lex-fridman-491.txtmarkdown158.5 KB
examples/steipete/data/transcripts/lex-fridman-clip-ai-agents.txtmarkdown25.9 KB
examples/steipete/data/transcripts/maximize-claude-code-subscription.txtmarkdown25.7 KB
examples/steipete/data/writing/2012-dont-call-willchangevalueforkey.htmlother59.8 KB
examples/steipete/data/writing/2012-hacking-block-support-into-uimenuitem.htmlother183.8 KB
examples/steipete/data/writing/2012-moving-on.htmlother62.2 KB
examples/steipete/data/writing/2012-nsurlcache-uses-a-disk-cache-as-of-ios5.htmlother54.5 KB
examples/steipete/data/writing/2012-pimping-recursivedescription.htmlother55.6 KB
examples/steipete/data/writing/2012-reboot.htmlother50.0 KB
examples/steipete/data/writing/2012-using-subscripting-with-xcode-4_4-and-ios-4_3.htmlother59.5 KB
examples/steipete/data/writing/2014-a-story-about-swizzling-the-right-way-and-touch-forwarding.htmlother112.5 KB
examples/steipete/data/writing/2025-applescript-cli-macos-complete-guide.htmlother110.7 KB
examples/steipete/data/writing/2025-automatic-observation-tracking-uikit-appkit.htmlother104.4 KB
examples/steipete/data/writing/2025-claude-code-anonymous.htmlother54.8 KB
examples/steipete/data/writing/2025-claude-code-is-my-computer.htmlother72.0 KB
examples/steipete/data/writing/2025-code-signing-and-notarization-sparkle-and-tears.htmlother101.6 KB
examples/steipete/data/writing/2025-commanding-your-claude-code-army.htmlother75.2 KB
examples/steipete/data/writing/2025-essential-reading-august-2025.htmlother78.7 KB
examples/steipete/data/writing/2025-essential-reading-july-2025.htmlother73.2 KB
examples/steipete/data/writing/2025-essential-reading.htmlother79.4 KB
examples/steipete/data/writing/2025-finding-my-spark-again.htmlother55.8 KB
examples/steipete/data/writing/2025-introducing-demark-html-to-markdown-in-swift.htmlother63.9 KB
examples/steipete/data/writing/2025-live-coding-session-building-arena.htmlother59.8 KB
examples/steipete/data/writing/2025-llm-codes-transform-developer-docs.htmlother59.9 KB
examples/steipete/data/writing/2025-logging-privacy-shenanigans.htmlother108.4 KB
examples/steipete/data/writing/2025-mcp-best-practices.htmlother69.0 KB
examples/steipete/data/writing/2025-migrating-700-tests-to-swift-testing.htmlother178.7 KB
examples/steipete/data/writing/2025-optimal-ai-development-workflow.htmlother67.8 KB
examples/steipete/data/writing/2025-peekaboo-2-freeing-the-cli-from-its-mcp-shackles.htmlother76.2 KB
examples/steipete/data/writing/2025-peekaboo-mcp-lightning-fast-macos-screenshots-for-ai-agents.htmlother67.1 KB
examples/steipete/data/writing/2025-poltergeist-ghost-keeps-builds-fresh.htmlother81.8 KB
examples/steipete/data/writing/2025-self-hosting-ai-models.htmlother114.9 KB
examples/steipete/data/writing/2025-shipping-at-inference-speed.htmlother70.1 KB
examples/steipete/data/writing/2025-showing-settings-from-macos-menu-bar-items.htmlother110.3 KB
examples/steipete/data/writing/2025-signature-flicker.htmlother69.5 KB
examples/steipete/data/writing/2025-startup-slop.htmlother69.7 KB
examples/steipete/data/writing/2025-stats-store-privacy-first-sparkle-analytics.htmlother58.6 KB
examples/steipete/data/writing/2025-stop-overthinking-ai-subscriptions.htmlother67.4 KB
examples/steipete/data/writing/2025-the-future-of-vibe-coding.htmlother88.9 KB
examples/steipete/data/writing/2025-understanding-codebases-with-ai-gemini-workflow.htmlother62.5 KB
examples/steipete/data/writing/2025-vibe-meter-2-claude-code-usage-calculation.htmlother71.3 KB
examples/steipete/data/writing/2025-vibe-meter-monitor-your-ai-costs.htmlother69.9 KB
examples/steipete/data/writing/2025-vibetunnel-first-anniversary.htmlother59.0 KB
examples/steipete/data/writing/2025-vibetunnel-turn-any-browser-into-your-mac-terminal.htmlother77.7 KB
examples/steipete/data/writing/2025-when-ai-meets-madness-peters-16-hour-days.htmlother70.4 KB
examples/steipete/data/writing/2026-openclaw.htmlother52.1 KB
examples/steipete/data/writing/command-your-claude-code-army-reloaded.htmlother59.4 KB
examples/steipete/data/writing/index.htmlother34.0 KB
examples/steipete/data/writing/just-one-more-prompt.htmlother59.7 KB
examples/steipete/data/writing/just-talk-to-it.htmlother80.8 KB
examples/steipete/data/x/curated-tweets.mdmarkdown42.1 KB
examples/steipete/data/x/openclaw.posts-original.jsonconfig68.4 KB
examples/steipete/data/x/openclaw.profile.jsonconfig478 B
examples/steipete/data/x/openclaw.replies.jsonconfig74.6 KB
examples/steipete/data/x/steipete.posts-original.jsonconfig95.5 KB
examples/steipete/data/x/steipete.profile.jsonconfig485 B
examples/steipete/data/x/steipete.replies.jsonconfig2 B
examples/steipete/examples/bad-outputs.mdmarkdown10.2 KB
examples/steipete/examples/good-outputs.mdmarkdown15.9 KB
examples/steipete/scripts/fetch-data.shexecutablescript11.4 KB
examples/steipete/scripts/pull-x.mjsscript3.9 KB
examples/steipete/scripts/weak-model-test.mjsscript17.7 KB
examples/steipete/tests/prediction-test.mdmarkdown13.2 KB
examples/steipete/tests/weak-model-results.mdmarkdown11.6 KB
examples/vitalik-buterin/MEMORY.mdmarkdown972 B
examples/vitalik-buterin/QUICKSTART.mdmarkdown3.3 KB
examples/vitalik-buterin/README.mdmarkdown6.9 KB

examples/steipete/data/transcripts/agentic-workflow-behind-openclaw.txt

View on GitHub
When you first start programming with AI
agents, you would say something like
that, please fix this or please build
that. But then as time goes, you try to
follow the tricks and tips on how to
optimize your workflow. They would tell
you to run eight AI agents
simultaneously, to have complex
orchestrations between multiple agents,
to chain agents together, to have custom
sub agents, to have a library of 18
different slash commands. After that,
you realize that the more you cut this
crap from the context and the more your
context is neat and organized and clean,
the better results you have. And then
you end up with something like this.
Hey, look at these files and then do
mediumsiz change. So it's not a large uh
full stack feature. So you kind of move
along this curve. At the beginning you
don't know anything. You do not have a
workflow. Then after that you start
complicating and over complicating
things. And then you come to a
realization that actually the simpler
the better. So this is actually my story
of working with AI agents for
programming and this graph is taken from
the blog of Peter Stainberger and if you
haven't heard of him he's the creator of
OpenClow or Cloudbot or Maltbot and he's
shipping like crazy. He's getting a lot
of things done and he's he's completed
projects more than companies and teams
would complete in years. In his blog he
said agentic engineering has become so
good that it now writes pretty much 100%
of my code. So the AI agent basically
writes 100% of his code. So how does he
do that to build such a viral product
like OpenClow just using AI agents.
Let's try to learn from him to see what
workflow he uses. So first let's start
with things to avoid. So I have
summarized the recent posts from his
blog. And here I give you the
conclusion. You do not need to elaborate
a lot on your prompts. So people online
like they're going to tell you to
basically have certain prompts for for
certain models. And they even try to
sell like prompt templates and this is
actually a scam. So never buy a prompt
template on the internet. First of all,
large language models are
undeterministic, meaning that if you ask
the same question with the exact same
wording, in some settings, it can give
you a completely different answer.
Moreover, the context also plays a role.
The context is not only the prompt. The
context can have the history of the
messages, your profile, some other
information that the LLM or the agent
has access to that is different from the
prompt. So this will also affect the
quality or the response. Moreover,
different models or different LLMs will
give you different results. So, it's not
reproducible. The idea here is that you
do not need to have elaborate prompts
anymore for agents because agents are
now very good in searching for context.
So, they're going to go to the to files.
They're going to search files in the
codebase and they're going to ask you
for any information that doesn't exist
in the codebase or doesn't exist in the
current context that the LLM has access
to. The second thing to avoid is sub
aents. So, the idea of sub aents is here
you have your main agent. So when the
main agent wants to perform a task, it's
going to spawn a separate agent and then
this sub agent will return the result
and then the whole context of how it
came into this this result will be wiped
out. First of all, okay, you isolate the
context but at the same time you lose
some parts of the context that might be
helpful in the overall task. This is one
thing. The other thing is that the
coding agents are going to spawn these
sub aents without your control. You
cannot know what's happening or what
context uh is basically being passed
between these sub aents. So you lose the
control over that. That's why he
suggests that whenever you want to uh
have some isolated task for example
let's say you want to search for
something and then you want you just
need to return the result but you do not
need the context of how this result has
been fetched. Uh you can just do that in
a separate uh in a separate agent. So
you have one agent for one task and then
you just open a new terminal and have
your new task in just a new u completely
new context. And then you can just take
the your findings and either apply to
the codebase directly where the first
agent is reading from or you can pass it
manually to the first agent. So yeah, so
this is this is the main idea. So the
third thing that he says to avoid is
wrapper tools. So wrapper tools
basically are like tools or software
that tries to wrap the existing coding
agents like load or or or codeex and
they try to give them extra capabilities
to vectorize the codebase and to save
the codebase in an index with the
current agent capabilities. They
actually are able to search the codebase
very very well. So you do not need to
have additional wrappers around these
LLM. So now what about code quality? So
if AI is going to write 100% of your
code then probably the code quality is
not going to be that great, right? So
here he gives some techniques or some
ideas on how to control the code quality
as much as possible. So first is the
blast radius thinking. So the blast
radius thinking is the following. Think
of your code base as this space of code
or this space of features and files. So
a feature for example lives in a subset
of these files. So it lives here in the
space. Another feature lives in another
subset of files in your codebase. And
here for example here is like a third
feature. The main idea is that you
shouldn't make changes or multiple
changes at the same time to regions of
your codebase that intersect, especially
when you're using parallel agents at
once. So this is agent one. This is
agent two. So agent one, for example,
can be working on this region. Agent two
could be working on this region. The
main thing is that they should be
independent, completely independent. If
they're not independent, it doesn't
work. Also another important idea is
that these changes or these regions
should be small enough for the agent to
be able to comprehend and to tackle
efficiently. So the larger the region
gets the more difficult for the agent to
maintain the quality that enables it to
implement that feature. So that's the
idea. Another thing is that he commits
directly to main. So he just has
basically a linear workflow. He doesn't
need like git work tree. And this tends
to be better than asking the agent to
work in multiple branches and then do
merges because this introduces merge
conflicts and additional overhead that
complicates a process. Okay. Now the
third thing are tools. So CLIs are
better than MCPS. The idea here is that
MCPs are chunks of text that live in the
context. So let's say this is your full
context. This MCP definition should live
inside your context. Another MCP also
lives here. Another MCP also lives here.
So they eat from the context. And also
it's unnecessary. For example, if the
model is not using any of these MCPS,
this data is going to still live in the
context. So, it's going to cause you
extra tokens and it's also going to
degrade the quality because it's just
occupying the context that could be used
for something else. So, the better idea
that he introduced is the use of CLI.
So, instead of having MCP's definitions
in the context, what you do instead is
that you just give it the name of the
CLI and what does it do? Then if the
model tries to use the CLI in some way,
it might get an error for example. And
once it get an error, it's going to go
and try to check the documentation of
that CLI and then it's going to come
back and use the correct command. So the
second thing is that he uses multiple
terminal with ghost ty. This is just
like a multiplexer for using terminals
in multiple windows. Actually I don't
use it. I just add a new terminal but
turned out that like a lot of people are
using it and uh he says that it's better
user experience when using this
terminal. Okay. Now the context. How
does he manage the context? As we said
brief prompts. This is one thing. The
second thing is he uses screenshots. He
just like screenshots for example the
result and he give it back to the agent
and LLMs now have decent visual
capabilities. So can they can read text
from the image and they can also analyze
how that fits to the implementation of
the current feature that we're building.
So the third thing is test same context.
This is very important when you
implement a feature ask the LLM or ask
the agent to immediately test it because
it has all the context it needs to test
this feature. So it's beneficial to
immediately test that and solve any
issues that it might uh encounter.
That's the idea here of the context.
Parallel workflows he basically have
multiple agents working in parallel
either in one codebase or multiple code
bases. Three to eight agents in a grid
scale by blast radius. Also atomic
commits atomic commits is also very
important. So this means that one commit
one clear change. Every commit has one
exact purpose which is implementing
something enhancing something solving
some bug and so on. And now the main
philosophy here or the main like uh
takeaway from this video is that just
talk to it, explore, develop intuitions
of how it works over time. Literally
like any other skill, just like you talk
to it and you try to improve over time.
Don't just like blindly listen to the
advice on the internet of how you should
have your setup. Experiment yourself.
Take those like tips from the people who
are successful. So this is a proven
workflow. And also uh as we discussed at
the beginning, less is more. So the less
complicated your workflow is, the better
the more you can get done and the better
your overall result will be. There is
actually one catch I want to explain but
before that let's check comparison code
versus codex. So he uses codeex. Why
codex? First of all it has more around
100,000 tokens in the context. It first
reads files before writing. So it tends
to explore the code base more thoroughly
before it does some change. The oneshot
accuracy is higher. So this is GP 5.2 2
and he's comparing it to clude opus 4.5.
It also handles large refactors better
than cloud opus 4.5. That's his
experience. So he has shipped a lot of
product and most of them are in
production like his trustable person.
The CLI itself is implemented in rust
which makes it faster and like more
smooth experience than clo code. Okay.
Now let's go to the part that most of us
won't like. So this is his token usage
on Codex. So he has used 285
billion tokens in total. To put it in
perspective, this is about 194,000 Harry
Potter series, 400,000 Lord of the
Rings, and 250,000 Bibles combined. So,
this is the amount of tokens he has
used. And the usage cost for that is
$51,000. So, most of us won't have this
enormous usage. But uh yeah, that's the
idea. The idea is that he keeps Codex
running for hours. So, he give it all
the context it needs, all the tests that
it needs to generate. the iteration
keeps running until the refactor is done
or until the feature is implemented.
It's very similar to the Ralph Wigum
thing, but it's just done in Codex. So,
if we just take a look at all of that,
we might be tempted to say that software
engineering is solved. Programming is
not needed anymore. But actually, there
are some still difficult parts and he
only is able to do that because of his
knowledge about programming and about
his extensive experience of what looks
good and what's not and how to steer the
model to come up with the best
solutions. For example, here he mentions
picking the right dependency and
framework to set on is something I
invest quite some time on. Is this
well-maintained? You want the agent to
use libraries that are well-maintained.
How about peer dependencies? Is it
popular? So, will it have enough world
knowledge so agents have an easy time?
Uh equally, system design. Will we
communicate via the websocket, HTML?
What do I put into the server and what
into the client? Often, these are things
that are a bit harder to explain to a
model and where research and thinking
pays off. There's still the human part.
There's still this experience of how the
system actually work. What are the
design decisions? For example, if you
don't know what he's talking about, you
will not even ask the question whether
to put some component on server versus
client for example. So if you don't know
that, you will just maybe after one two
weeks get some like performance
degradation, you will not even
understand why this happened to begin
with. But if you know from the
beginning, you can just design and
architect the system in a much better
way. And uh yeah, so that's what he's
talking about like to do the to do the
research properly and the system design
properly will pay off in the long run.
So yeah, that was all. I I hope you
learned from it. Thank you for watching.
See you in the next one. Bye-bye.
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: 198kUpdated Claude CodeCodexVerifiedInstall
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: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
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: 194kUpdated Claude CodeCodexVerifiedInstall