Seltzer Genius 🥤🧠

imo its important to completely and utterly understand that everything everyone ever says is motivated by emotion you can put whatever horseshit on top of it you want, if i know how you feel i know why you built it

@RhysSullivan interesting pattern--i dipped to my own workflows and ghostty when you changed the billing model but it looks like you ran with it. will check out this week

executor now has a desktop app! add whatever MCPs / OpenAPIs / GraphQL servers you want once and then every agent can use them converts them all into code mode under the hood, so you can have thousands of tools and no context bloat everything stays 100% local on your device

@doodlestein @osekkat is there value in having codex based agents use goal mode in conjunction with bead execution? might be unnnecessary once you are at the level of granularity and plan hardness required to cut beads to begin with also thought of using it for "research completeness". havent tested

@osekkat No, I am using them separately for now in a sort of complementary way. For big initiatives I’m still using ntm mostly. The goal mode is nice to allocate one long session to a single project to do lots of housekeeping type stuff versus for implementing a ton of new beads.

The combination of Codex goal mode, my ntm orchestration tool, and all my new "super skills" (like /reality-check-for-project and /simplify-and-refactor-code-isomorphically) have led to another explosion in my useful output just in the last week or two. The acceleration is nuts!

@doodlestein hes right

The bitter lesson applied to programming today is that you need to find ways to productively leverage as many tokens as possible. People roll their eyes at this concept because it's easy for this goal to be misapplied and lead to wasteful nonsense. But if you just think about it, it's obviously correct. The models are VERY smart now and can work basically 24/7 for you, and you can conjure up as many of them as you want. The only limitation is the number of Max/Pro accounts you have, and controlling enough fast machines with a lot of RAM and CPU cores to run them all. But that's just what you need to be in a position to do this. The bigger impediment for most is on the management side: how can you usefully direct and harness so many agents? And the answer is that you need to work towards the goal of removing yourself from the equation as the bottleneck for as many projects as possible so that you can be making real progress on them every day. This frees you up to focus your energy on the couple of things that are most critical for any given project and so you can do the planning and conceptual work on new projects without having everything else come to a halt. The solution to doing this in practice is to use the right tooling (like Agent Mail, beads_rust, bv, ntm, cass, dcg, rch, etc.), the right planning methodology (the complete Flywheel guide), and the right skills (from jeffreys-skills.md). Are there other ways to go about this? Sure, everyone is increasingly converging on the same basic principles. But I know for a fact that: * my tools (which are 100% free and open-source) really work (agent-flywheel.com/tldr); * my planning methodology and workflows really work (I've already used them to make extremely powerful ground-up software projects like asupersync and FrankenTUI; see agent-flywheel.com/complete-guide); * my skills enable a different level of autonomy and economy of expression (I've seen a huge inflection in my own output, which was already pretty crazy before, since I started using them universally in all my work). So that's why I recommend them to people. There are nearly 1,000 people in the Agent Flywheel Discord (discord.gg/gnCHsYDR25) now who can tell you the same thing: this approach really works, it's not just hype. I'm not trying to justify some crazy VC valuation. It's just me cranking away with my 52 AI subscriptions and giving 99% of it away for free. The only things I directly monetize are my skills, and they are just $20/month because I'm trying to prioritize growth (hey, it works for Netflix and Spotify!) So give it a try. If you already have Claude Max and/or GPT Pro, you have most of what you need to get started. You can run the tools on a mac, or better yet, rent a decent cloud server for $50 to $100/month, which works better and can scale more. The wizard on my agent-flywheel.com site walks you through every step of the process. I promise you that you won't regret taking the plunge. It's a lot to digest and understand and apply, but the rewards are magnificent right now because there's still huge alpha in being able to do this while 95%+ of the world thinks it's impossible or a pipe dream.

@namcios this actually sucks ass if you care about determinism in your workflows definitely keep telling people this shit though so i keep my edge for a while longer

Dois engenheiros da Anthropic acabaram de mudar a forma como devs pensam sobre IA. Barry Zhang e Mahesh Murag subiram no palco do AI Engineer Code Summit e disseram uma frase que incomodou muita gente: "Parem de construir agentes. Construam Skills." Em 16 minutos eles provam que a indústria inteira está resolvendo o problema errado. Aqui está o que a maioria não entendeu: → Skills são pastas. Literalmente pastas com arquivos markdown. → Elas ensinam ao Claude o SEU fluxo de trabalho, a SUA expertise, o SEU domínio. → Um único agente genérico + biblioteca de Skills específicas supera dezenas de agentes especializados. → Fortune 100s já estão deployando Skills em escala pra ensinar agentes sobre processos internos. → Times de produtividade com 10.000+ devs usam Skills pra padronizar como código é escrito. A analogia que eles usaram é perfeita: Quem você quer fazendo seu imposto de renda? O gênio com QI 300 que nunca viu legislação tributária, ou o contador experiente que faz isso há 20 anos? Inteligência sem expertise é entretenimento. Expertise empacotada é produtividade. O que mudou: a Anthropic parou de tentar criar agentes diferentes pra cada domínio. Perceberam que com Claude Code, o padrão é sempre o mesmo. Um modelo acoplado a um runtime com filesystem. A diferença entre um agente medíocre e um extraordinário não é o modelo. É o conhecimento de domínio que você alimenta. Skills resolvem isso com progressive disclosure. O agente só carrega o nome e descrição da skill. Quando relevante, puxa o SKILL.md. Quando precisa de mais, navega os arquivos de referência. Zero desperdício de contexto. Isso não é uma feature. É uma mudança de paradigma. Quem entender isso agora vai operar em outro nível daqui a 90 dias. Quem ignorar vai continuar escrevendo prompts de mil palavras toda vez que abrir o chat. E ainda explicar de novo e de novo o que “realmente” quer.

@doodlestein @garrytan put a bunch of codex agents on it impersonating users after doing qa until they find bugs to report and you dont even need the users anymore

You should try automating the PR review process and also the GitHub issues review process. I’ve done that for all my projects using custom skills and my repo_updater tool, and it lets me get through many dozens of them every day in just a few minutes of my human time. It only surfaces things for review that actually require real judgment. If it’s a confirmed bug, it just fixes it. If it’s an obviously good feature that fits into the project well, it just implements it. And it responds to people using the gh utility and just manages the whole process nicely. I have them on my site (see images below), but they’re pretty customized for the way I do things (I don’t merge any outside PRs but do mine them for ideas). But you could easily customize them or just implement something similar yourself. It’s definitely worth spending the time on. I used to worry that people would be offended getting responses from an agent impersonating me, but it turns out that people just want their issues resolved quickly and don’t seem to mind that aspect of it.

I didn't drink at my birthday party so I could start crushing PR's all day today. I have 99 PR's to work through from the community, but a hangover aint one First big feature of the day: a chief security officer to audit your attack surface (informational only, please hire a pentester for real production code and mission critical situations!)

@trq212 this is a very targetted and sorely needed tool as those of us who built insane confabulations of orchestrators collapse back down into claude code as expected as it becomes capable of all of the things we orchestrate natively really good product insight here imo

We just added /btw to Claude Code! Use it to have side chain conversations while Claude is working.

"Hedonism during the decline of the Roman Empire is often characterized as a period of intense elite decadence, featuring extreme luxury, sexual excess, and indulgent banquets, particularly depicted in criticisms of the later empire."

@zeeg working on something for this...

Spent the last few days on Codex. 5.3 is a great model but its explanations need dumbed down for me to survive the context switching 😅

@hopes_revenge lotta catching up to do

probably gonna be talking to my computer more . maybe moaning

there was a time five or more years ago where someone sending you a comprehensive breakdown in response to a business problem was super impressive and demonstrated how lucidly they understood the situation now it almost always means the exact opposite, and the receiver knows it

this example is so funny because it assumes the receiving party will be reading that info dump the reality is the receiving party will immediately fog over when they see it's AI generated then throw it through their Claude cowork sure you can see where that's going

@keysmashbandit I feel similarly. It would be a miracle and a boon for humanity if not for capitalism, and maybe it will be a wedge to that end yet. Or maybe it's the dawn of a turbo dystopia I will regret bringing a child into.

I understand some dooming, but at the same time I feel so incredibly lucky to be alive to witness the birth of intelligent machines, and to even be allowed to play with them and explore many things about them. And that is my primary feeling regarding the situation

@billybrizzle420 you are seen you are valid

We’re here, we’re not queer, but we’re close

I’ve had enough of straight male bottom erasure

@somefoundersalt good for them

fun little toy. I did one with trading bots connected to opencode so their could modify their data sources and write analytical tools etc spoiler alert it didn't make them awesome at trading it made them technical wonks in a perpetual and expensive podcast episode lol

everyone did this a year ago

@RhysSullivan @opencode congrats dude great product

@tenobrus idk about langchain dog what replaces that

gigafucked: - grammarly - calendly - miro - retool - webflow - langchain - writer - harvey - glean - expedia - monday fucked: - accenture - intuit - notion - jasper - canva - alphasense - postman - airtable - talkdesk - sierra - zapier - replit - solace probably fucked: - cursor - pilot - clay - mercor naively seems fucked but so competent / plugged in they seem to be figuring it out on the fly anyway: - linear

@bpodgursky no youre actually behind

This is maybe obvious but I think a lot of the reason people write complicated CLAUDE md files, skills, fancy custom wrappers is not because they meaningfully help but because they're unsettled and grasping for ways to add value to a dev workflow where the defaults work fine.

@Hesamation no this is super stupid and they should be embarassed

everyone has been dunking on her, “ShE’s tHe HeaD of SafeTy At Meta” or “wHy did She PoSt tHis?” let’s talk about this: > doing research in safety doesn’t mean you’re incapable of making mistakes > 1000s of OpenClaw users made the exact same mistake as her. > she was the only victim, it wasn’t a corporate mistake affecting users > maybe we should talk about the preachers who don’t warn enough about it? if you talk about making billions with OpenClaw maybe you should also share what you did to make sure it’s secure. the most stupid take on this was “why did she openly share this? it’s bad PR” well if you’re a safety and alignment lead and have AI expert connections on social media, this is exactly what you need to post and raise awareness about, not your dinner pics, even if you take the hit for it.