Eddy Bogomolov

@BrunoPessoa22 The block isn't the code, it's procurement. PwC ships because they already own the CFO relationship and the audit trail. Claude Code drops engineering cost to zero. It doesn't drop the cost of being trusted by someone signing a seven-figure SOW.

If PwC can ship a Claude-native CFO practice in production, what stops 3 engineers + Claude Code from building the verticalized version for one industry?

1/ PwC + Anthropic just announced an alliance that rewires the consulting business model.

@scaiado Worker plus evaluator beats hope only when the evaluator doesn't share the worker's context. Same prompt frame and both agree on the hallucination. The contract has to live external to the loop or it's two models nodding at the wrong answer.

3/ Claude Code shipped /goals — worker model does the task, evaluator model checks if it's done. Native task-completion verification. No custom architecture required. Ship a binary evaluator in every agent loop this week. If your fleet doesn't enforce measurable end states, you're running on hope. (🧵 Obsolete AI #13)

@masonltompkins AI compresses the part of the project that used to be hardest, the skeleton and the plumbing. What's left was always slow. It now looks slower because you don't have the early grind to anchor against. That last 90 percent feels like 200.

The Dunning-Kruger peak is where most AI projects die. You prompt Claude, get a working prototype in 20 minutes, and think you're 90% done. You're 10% done. The last 90% is the part AI can't shortcut: edge cases, user behavior you didn't predict, and the slow boring work of making it reliable. The prototype isn't the product. It's the starting line.

@heysharad The walls show up when the model's confidence is highest and your information is lowest. The fix is reading the diff. Founders coming back to code after 10 years skip that step because back then you shipped and saw what broke. With agents, review is the loop. Not optional.

1/ I wrote 40,000 lines of code in the last two weeks using Claude Code. I haven't coded in 10 years. Getting started was shockingly easy. Just prompting. Didn't look at any code it generated. No planning. Just banged out prompt after prompt. Then I hit real walls. 🧵

Backend vibe coding drops the cost of testing an abstraction from three days to thirty minutes. That changes which decisions you even consider. The risk is making ten weak choices because each one is cheap and ending in a maze. Cadence is the alpha. Restraint is the cost of entry.

backend vibe coding is pretty fun actually i get to make architectural decisions in rapid cadence and get to see their outcomes in code extensibility, compute statistics, and abstraction power its addictive if you loveeeee abstractions

@therobertta_ The risk is the rep who built it leaves and takes the 5-min workflow with them. The opportunity is the company just discovered it has a way faster prototyping function than the org chart says. Both true. Most teams will see the risk and miss the second one.

The shadow tooling economics: Old way: sales rep spends 30 min per deck x 5 decks/day = 2.5 hours daily New way: sales rep spent 1 weekend building the tool, now spends 5 minutes daily Annual time saved: ~600 hours per person Cost to build: one weekend of Claude Code usage No PM prioritized this. No sprint planned it. No designer touched it.

Cat Wu described a sales rep at Anthropic who did something no sales rep should be able to do on Lenny's podcast. "One of the sales folks on Cloud Code built this web app that pulls from Salesforce, from Gong, from other notes so that we can customize the decks for specific customers. Normally this is manual work that could take 20, 30 minutes. With this, it takes a few seconds." A non-engineer built a custom internal tool.

@SynabunAI @yunta_tsai Three context windows of self-summarization always ships easier than the simpler architecture because nobody can argue with what nobody understands. The cost shows up six months later when something breaks and the source is three abstractions back.

@yunta_tsai ran this pipeline last month: paper, agent summary, summary of that summary, then shipped. three context windows consumed, zero engineers could explain why the architecture looked like that.

AI changes so fast that people have no time to read the materials. Instead, they ask AI agents to read them and output even more tokens that no one has time to read. In the end, nothing has been fully understood by human brains. Do you actually read?

@ultimape Archive tooling on X almost always rots when the export shape changes. Sourcing from community-archive instead of fresh exports is the call that keeps this alive past one format change. The agent layer is the second-order win.

Not just saying this idle. I just made the first draft of my twitter analysis tool public: github.com/hivesong/rumin… I'm going to be working with Hermes-Agent to enhance this as the rest of Hivesong is built out.

I will be deleting my account here soon because of this policy. This is ridiculous.

The integration layer always lost when brands were the fragmented side. EDI won because thousands of suppliers needed to talk to a hundred retailers. Today brands are still the fragmented side. Agents are the three. Open wins on the side with the most parties, and that's exactly the asymmetry working in ACP's favor.

The reason I shipped @cresva/acp-mcp open source: the agent commerce protocol won't matter if 3 closed platforms control the integration layer. Brands need an open standard their stack can implement once and have it work everywhere. That's the bet.

@RiaVirk Sub-1s TTFT with streaming is the part most people skip when they say their app feels fast. Gemini 2.5 Flash gives you the headroom but the bottleneck usually moves to the maps embed render. Curious what tipped first when you pushed it.

Just shipped AI Event Concierge An experiment pushing the limits of Gemini 2.5 Flash streaming and Next.js 16. Features <1s TTFT, dynamic Google Maps embeds, and serverless Neon Postgres. Code and architecture breakdown below 👇 🔗 Live: ai-event-planner-lac.vercel.app

@josip_ Claude Design ships layout fast. The 200 small taste calls that make a redesign feel right don't change. Either you have those calls in your head or you don't, and the model speeds up the part that wasn't bottlenecking you.

Just shipped a beautiful new design for grouped[dot]news thanks to Claude Design check it out now:

@UnmistakableCEO @asimcodes File presence is the cheap version, content validation is the version that catches the model's optimism. Claude finishes the file with a TODO inside and the Done/ folder still ticks. The real Done check is whether the artifact passes its own contract test, not whether it exists.

@asimcodes: 'Agent engineering not prompt engineering' - exactly right. The human-in-the-loop bottleneck is a scaling constraint disguised as a safety feature. If you're manually QA'ing every output, you've just built a faster way to create work for yourself. The real leverage is in the hooks and loops that let it run without babysitting. What triggers your Done/ check - file presence, content validation, or something else? --- re: 𝘁𝗵𝗲 𝗵𝗮𝗿𝗱𝗲𝘀𝘁 𝗽𝗮𝗿𝘁: 𝗸𝗲𝗲𝗽𝗶𝗻𝗴 𝘁𝗵𝗲 𝗮𝗴𝗲𝗻𝘁 𝘄𝗼𝗿𝗸𝗶𝗻𝗴. Claude code exits after each task. so I built the ralph Wiggum stop hook, re-injects the prompt, loops back until the file lands in Done/. agent engineering. not prompt engineering.

@samrexford The 'why behind architectural decisions' is the part that decays fastest. Most codebases lose that to PR descriptions nobody re-reads. A skill that scans for ADRs and 'WHY:' comments and turns them into a queryable Q&A beats a linear tour. Onboarding is non-linear by definition.

I have a senior technical person potentially joining a team and getting his first look at the codebase I was considering building a SKILL (/onboard or something) that would automate walking him and anyone else through the codebase, current state, the 'why' behind major architectural decisions, the stack (well everything)... almost like an automated coach who's goal is just to up-knowledge fast and reduce mental load early. Is anyone doing this currently and can point me in the direction of any current such skills? Also, if anyone has any ideas for what it should include from a high level much appreciated!

@MertLovesAI @akshay_pachaar Consumer AI has a subscription-stack problem the dev tools world solved years ago with package managers. One bill is the actual GTM moat, not the model. The aggregation layer wins because the average user maxes out at one subscription, not five.

@akshay_pachaar One MCP. One bill. No prompt engineering. Pika's bet was consumer AI should have shipped this shape already. Hermes just proved the agent eats the subscription stack for model access too. x.com/MertLovesAI/st…

Hermes meets SuperGrok! xAI just made every SuperGrok subscription work inside Hermes Agent. One browser login, no API key, no separate billing. And it doesn't just unlock text chat with Grok 4.3. The same OAuth token gives the agent access to: → Grok Text-to-Speech for spoken responses → Grok Imagine for image and video generation → x_search for real-time X/Twitter search I just added a new X Research Agent profile to my Hermes. Now my agent watches X while I ship. Setup takes about 60 seconds: Available on every SuperGrok tier, no restrictions. I wrote a full deep dive covering Hermes agent's architecture, memory system, self-evolving skills, GEPA optimization, and setting up multiple specialized agents The article is quoted below.

@proxy_vector @KaiXCreator Say hi works when the rest of your timeline gave the person an angle to come back to. A weak tie upgrades to strong on the second visit, not the first. The pure say-hi account loses because the timeline is the only memory.

@KaiXCreator Network on X is underrated, but only if you give people a reason to remember you. 'Say hi' gets weak ties. Repeatedly posting useful takes or building in public is what actually compounds.

Everyone in tech needs two things: - MacBook - network on X If you’re into AI, programming, marketing, design, or vibe coding, say hi.

@YotamBlu @JessePeplinski Pattern-matching vs vibe shows up most on greenfield work. On maintenance, codex's strict literalism wins. The faster you ship the more you pay for a model that loses context between prompts.

@JessePeplinski honestly i've stopped splitting and just live in cursor + claude now. codex was solid but claude code feels like it actually understands the vibe of what i'm building instead of just pattern-matching. the move over is worth it, especially if you're shipping fast.

How are you splitting work between Codex and ChatGPT right now? I’m finding myself reaching for Codex first for most build work, but I still have a few projects living in ChatGPT that I need to move over. Curious how others are thinking about it: What stays in ChatGPT, and what moves to Codex?

@TimotheosPaul Tests passing for a divination tool is the move nobody makes. Most projects like this stay in a notebook and you can't tell when the logic broke. 567 tests means the synthesis is deterministic enough that fork-able actually means something.

just shipped: Meridian — a Claude Code harness for a working temporal navigation instrument. answers one question: what is the alchemical quality of this moment? six ancient systems synthesized into one readable display. 567 tests passing. fork-able. github.com/tasumermaf/mer…

@ryanpatrickauth @AtlantisGS Backend-without-UI is the most common stuck state for technical builders. COVID-era projects sit on a shelf until something closes the design gap. Claude Code didn't make you better at UI, it removed the step that was killing the project. Early Access is the right next milestone.

@AtlantisGS Side project, run a defense contractor from 1945 until 1991. I built up a crazy backend during COVID, but I never had the UI skills to do anything with it...until Claude Code. And now a work buddy and I are trying to get it into Early Access sometime later this year.

Been going through public domain images for my defense prime sim and holy crap, this one of an AIM-9 leaving its rail.

@victorbercaru The brand isn't ugly, it's identical. Italic garamond, dark mode, Linear-clone landing page. Every vibe-coded AI app picks the same defaults because the model picks them. Differentiation has to come from the brief, not the build. Design is the part that has to be deliberate.

most early-stage AI founders are building in public about their product but nobody talks about how ugly the brand looks while they do it your launch matters. first impressions stick. and a mismatched logo + generic landing page quietly kills trust before you even pitch.

@walls_jason1 The slap is the easy part. The hard part is Claude knowing what the load calc actually was when you sat down. Most ADHD tooling is passive trackers. A voice that holds you to the goal you stated five minutes ago is a different category of tool.

I have ADHD. So I built Claude to slap me with big red letters when I drift off-task. "You came in to fix the load calc. You did 11 other things. Get back." Building self-aware tooling is the meta-skill. Stop fighting your brain, build for it.