Assaf Kipnis

I used Claude to make a demo video for KTLYST All I gave it was my 1 pager and the link to the tool. I'm floored

Today's episode in five stories. Apple's reportedly planning an M7 Ultra with up to 1.5 TB of unified memory. Frontier-scale open weights held entirely in memory on one desktop. The local inference math flips. reddit.com/r/LocalLLaMA/c… The one I can't shake is J-Wash, a way to customize open-weight models built on Anthropic's Jacobian-Lens work. The author's word for it: brainwash. Personalizing a model and rewriting its behavior are the same move. Community fine-tunes are a supply chain. reddit.com/r/LocalLLaMA/c… Simon Willison found the coding-agent effect in Datasette's code-frequency chart. A real before-and-after from someone who ships constantly. #atom-everything" target="_blank" rel="nofollow noopener">simonwillison.net/2026/Jul/13/da… Matt Pocock shipped the real skills directory behind a daily Claude Code workflow. Not a tutorial. A live config. github.com/mattpocock/ski… gstack is Garry Tan's Claude Code setup as 23 opinionated tools, CEO to QA. A one-person company as agent roles. github.com/garrytan/gstack Full episode: podcast.ktlystlabs.com/api/episode?d=… h/t @mattpocockuk @garrytan

Everyone says Fable gets "lost" in long sessions. That it gaslights you, and the fix is to keep it humble by having it "spar" with GPT. It's not the model. It's where you're keeping the memory. If the state lives in the session, the context fills up and the model loses track of where it stands. Then it fills the gaps with confidence. It didn't lie. It forgot, and it doesn't know it forgot. Policing that with a second model treats the symptom. Take memory out of the model instead. Write the plan and the current status to a file. When it drifts, reset it and reload the file. Don't argue with it. I run agents that build for hours without me watching. They don't get lost. Not because the model is humble, but because the memory isn't in the chat. It's on disk. Every session starts by reading the file and ends by writing it back. Long sessions aren't the problem. Trusting the context window to remember is.

Today's episode in 3 stories. Anthropic pushed the Fable 5 cutoff back again and kept Claude Code's raised weekly limits in place. GPT-5.6 Sol is at Fable class now, so the best model stays in more hands. That's competition doing its job. #atom-everything" target="_blank" rel="nofollow noopener">simonwillison.net/2026/Jul/12/bu… ECC packages a full production harness for coding agents, with 67 specialized subagents and a security scanner called AgentShield underneath. 229k stars. People stopped hand-rolling their setups. github.com/affaan-m/ECC shot-scraper 1.11 gives your agent eyes. Scriptable screenshots from the CLI, and one browser session now serves video plus multi-shot runs. Quiet workhorse. #atom-everything" target="_blank" rel="nofollow noopener">simonwillison.net/2026/Jul/12/sh… Full episode: podcast.ktlystlabs.com/api/episode?d=… h/t @affaanmustafa @simonw

Have you heard of the Claude /doctor command? I ran it today - It found 10 plugins I hadn't opened in 8 days and shut them off. That's the boring part. Agent tooling piles up cruft you never see. Here's the part I didn't expect. Mid-cleanup it broke. A bad flag ate the filename and zeroed my settings file. Then it caught itself, grabbed the backup it had made seconds earlier, and put everything back. The cleanup wasn't the interesting part. The self-recovery was. Most tools fail and hand you the mess. This one failed, saw it, fixed it.

My ADHD makes it so I can't be consistent with GTM flows. I like building and I get stuck there. So I gave a Claude agent full control as a GTM manager with a fleet of agent employees. I named him Cole and he is my employee. Cole's job is to make me money.

Everyone is shipping "context layer" tools for AI agents this month. I built a system to test the hype instead of retweeting it. It scores every tool my daily feed surfaces against my own stack. Not "is this cool." Does it close a real gap I have, and does it beat what I already run. First category up: context-layer MCPs. The pitch is everywhere. Your coding agent relearns your repo every session, so bolt on a graph that remembers. Four of them showed up in one week. That much convergence usually means a real gap. So I evaluated all four. One made the loudest claims. Its engine is closed source and it phones home a leaderboard. The loud number is the tell. Off the list. Two were clean. Local, open, read-only on your code. Worth a real test. The fourth wasn't software at all. Markdown templates and a folder convention. I read the whole thing in ten minutes. That's where the lesson was. It listed five failure modes: zero memory on restart, handoff files clobbering each other, memory truncation, lossy agent handoffs, decision drift. I had already fixed four. Memory topic-files, an append-only learning log, a decision log with origin tags. Built months ago because I watched teams lose institutional memory at Google, Meta, LinkedIn. Same failure, different building. But it caught one I missed. My session handoff is a single shared file. Two terminals wrapping at once, last write wins, one session's context gone. A race condition in my own tooling. Fixing it now. I didn't adopt a single tool. The exercise still paid for itself. The tool you reject teaches you as much as the one you adopt. It tells you what you already got right, and shows you the one thing you got wrong. What's the last tool you didn't buy that taught you something?

We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.

Today's episode, all of it. vLLM v0.25.0 makes Model Runner V2 the default for all dense models. 558 commits. If you self-host open models, your baseline just moved. github.com/vllm-project/v… Claude Code v2.1.207 turns Auto mode ON by default for Bedrock / Vertex / Foundry. Autonomous coding agents in enterprise clouds, opt-out instead of opt-in. That's a governance shift hiding in a changelog. github.com/anthropics/cla… Who does Anubis actually stop? Serious AI crawlers pay the proof-of-work cost easily. The wall filters the polite bots and little else. fzakaria.com/2026/07/09/who… Tools: Graphify (82k stars): folder of code + schemas + docs becomes one queryable knowledge graph for your agent github.com/Graphify-Labs/… agent-skills by Addy Osmani (77k stars): senior-engineer discipline you install instead of prompt github.com/addyosmani/age… open-design: local-first design engine, your coding agent does the work, real files out github.com/nexu-io/open-d… rag-rat: MCP server that gives your agent the history behind your code, so it doesn't relearn the repo every session reddit.com/r/mcp/comments… codex-ollama-proxy: full Codex tool stack against models you run yourself reddit.com/r/mcp/comments… Full episode: podcast.ktlystlabs.com/api/episode?d=… h/t @graphifyy @addyosmani

@AmSmoothieWhat Exactly. Initially I was very much against it, but I created pretty robust guardrails and now I feel more comfortable with it

@Kipnis_a Early experiment on scaling personal productivity via full AI delegation.

being a “full time content creator” seems less like having a job & more like running a small business where the inventory is your personality & the warehouse is your prefrontal cortex. idk how anyone does it for more than few years without burning the hell out.

@AmSmoothieWhat Right now I got it almost totally automated where Claude writes, posts, fixes my sites, sends dms and responds. I'm waiting to see how that produces

@Kipnis_a Got it. What's next?

I have about 20 different Claude instances that perform multiple functions. I noticed that I keep learning the same lessons every time on every instance. So I built a mechanis that fleshes out lessons that I learned when using each one. And then a tool aggregates all of of them every morning and teaches the lessons to all of the other instances.

New eps every day on podcast.ktlystlabs.com

Today's episode, four things worth your time. OpenAI retired Atlas. The browser moves inside the ChatGPT app and Simon Willison calls it the end of the AI-browser category. He's right, an agent doesn't belong in the browser that holds your logged-in sessions. x.com/simonw/status/… GPT-5.6 dropped and picking the right model is now the hard part. Early read: Sol on Medium is the new coding default, replacing 5.5 at xhigh. x.com/simonw/status/… LWN on AI scrapers hammering open-source infra. A grounded report from the operators keeping sites alive under crawler load. lwn.net/SubscriberLink… Ollama 0.32 adds Qwen3.5 and Qwen3.5-Next for local runs. One pull and the newest Qwen line works offline. Still an RC. github.com/ollama/ollama/… Built by @ollama

Ollama 0.32.0 adds Qwen3.5 and Qwen3.5-Next support. The newest Qwen generation, running locally with one pull, no hand-wired parser. Still an RC. Fine for a dev box today, but I'd wait for the stable tag before anything scheduled touches it. github.com/ollama/ollama/… Built by @ollama

@UK_Daniel_Card You should be 99% of those claims are BS

I think LLMs (AI) are really fucking cool/useful and will help us change how we use technology! but.... also..... i am very skeptical about lots of the 'cyber' claims....

Honest question to the people posting the giant Obsidian graphs and saying "Look at my AI brain" What does that giant graph actually give you besides a pretty picture?

@mahdinchoudhury yeah, after 6 months of fighting with it, I just built my own

Notion is so messy and overwhelming that I literally never want to use it.

@DanielLockyer Talk to any Meta Eng

Is GraphQL still a thing? Haven't heard about it for years

@plainionist I would ensure that I am tracking every PRD and spec from start to finish and incorporating former RCA learnings in each one. Also I would have the AI review the plan with fresh eyes every time

Serious question: If you started your agentic dev setup from scratch today, what is the ONE thing you would change? 🤔

Claude should just show your Current Session Usage in the bottom left corner of the app. Kind of ridiculous we constantly have to navigate and do two mouse clicks to check it manually. @DarioAmodei plz fix