Awww, Sol is very polite today.

When people debate whether human should review every single line of code, I get the essence of it. But then I see stuff like this IRL: Had I not looked at the code, how would I have caught this bad code? Btw, it's not about a specific harness or model. It's how things evolve in a context window.

I love that Codex (or now ChatGPT Codex 🙄) has this feature to annotate any text in the response and ask questions inline. I can keep adding a bunch of annotated questions and send them all in one prompt. It is easy, intuitive and efficient. Codex gets extra credits, and Claude Code gets none.

We've opened our free and self-paced Agentic Coding workshop. Build a real-time, multi-user Idea Board with Claude Code - deployed to Vercel on your own URL in 3-4 hrs. AI foundations + 9 hands-on steps. For anyone who wants to understand how agents actually behave. latentsignal.org/workshop

TIL: Google's NotebookLM is a great tool to get a deep understanding of anything on the internet. While reading this article from HN injuly.in/blog/napkin-..., I wanted to know how the math works. I gave it to NotebookLM with specific questions, it explained well and gave me a good understanding.

100% agree.i signed up and it is refreshing to see a simple service with no frills. It just works. @pushover.net 👏

The old rule: "Show don't tell." Me to the coding agent: "Tell, don't show (off)" When I ask a question, just answer it. don't try to rewrite the code. I don't know why they are so eager to change everything. Too much bias to action!

We posted Memento in Hacker News: news.ycombinator.com/item?id=4855... Join the conversation and tell us what you think. We would love to hear your feedback.

A completely practical setup to run open-weights model locally on my M2 Max 64 GB: - Run llama.cpp UI using llama-server `./build/bin/llama-server` - use Unsloth's Gemma-4 26B-A4B model `unsloth/gemma-4-26B-A4B-it-GGUF:Q4_K_XL` Ofc this setup works great for knowledge synthesis, not exploration.

AI is a tool, there are immense benefits, big risks too. We use many AI tools everyday and have a measured take on where it shines, where is fails and where it is headed. Latent Signal is the channel through which we share our learnings and ideas. We hope to share everyday from now.

Claude these days: 🙄 Me: "Keep the reset command separate and make it failure-proof." Claude: "You're right on both counts. Let me reset (heh) and answer." Didn't even bother to say pun intended. Just "heh" and move on!

I am loving the side chat feature in Codex. You can... 1. Select any text in the main chat window and choose 'Ask to side chat'. 2. The side panel opens and shows the selected text. Ask your question here. 3. You can select a (smaller) different model other than main chat.

Images created by the new image model announced by Microsoft today - MAI-Image-2.5 Prompts: 1. “create an image of an iceberg in a beautiful ocean with stunning photorealistic way” 2. “add a bluewhale jumping from the water near the iceberg 3. "show the sun setting in the horizon"

Did you know that civil society organizations are eligible to apply for Cloudflare's Project Galileo? We provide free cybersecurity protection against DDoS and other cyberattacks targeting public interest groups. Learn more: https://cfl.re/3IgTAch

I'm loving the new Copilot GUI app (early access). Very professional, clean, and intuitive UI. Things are exactly where you'd expect. Feels like a well-engineered, production-grade app. Unlike some popular coding agents that still feel hacked together with a disjointed UX. Great job, @github.com 👏💯

Cloudflare's security team spent the last few weeks testing Anthropic's Mythos against fifty of our own repositories. What we learned about offensive AI, why faster patching is the wrong reaction, and what the architecture around vulnerabilities has to look like next. https://cfl.re/4eSYw7W

I am not going to name names, but... I asked my coding AI agent to “improve” my simple parsing code that extract a folder path from inline markdown text. It responded by adding FOUR new dependencies just for that job. Me: "Bro!That’s too much." Agent: "Fair pushback." Me: (facepalm) Revert it! 😡

most of these github actions driven breaches are because of pull_request_target. here's some tips of what to look for when reviewing your own workflows. worth noting - it is safe, and necessary, when used correctly 🧵

Found something useful today. The pi agent couldn’t read this webpage, so I used the Obsidian Web Clipper to copy it as Markdown, and pasted it into the chat. Worked immediately. pi was happy! Small win, but a sweet one.🎉 Wonder if this can be built as a skill with Obsidian CLI - @stephango.com

I finally completed the migration of my personal website: annjose.com It started as a Hugo to Astro migration and became a full redesign and a real agentic coding project. 11 years of blog posts, images, comments, tags, search, OG images, deployed to Cloudflare. annjose.com/blog/hugo-to...

All the demos shown in the Code with Claude event is using Opus 4.1 with 1M context X-high effort.

Everybody say Thank you Claude! 🎉 Anthropic at Code with Claude event - we are increasing rate limits on Claude Code - doubling the 5 hour limit for Pro, Max and Enterprise plans! 🔥

Me to all the coding agents (Claude / Codex / pi /...), Tell me what you are going to do, Do that and nothing else, Tell me what you did. Thank you!

Gemma-4 E4B running on my Google Pixel 8 phone. Impressive audio transcription, translation, image recognition. Knowledge retrieval not so much (understandably as small model).💯 The model is a lot more capable than gemma-3. It's time to update the post I wrote last year! annjose.com/post/mobile-...

Nextdoor notification settings: - 8 topic groups - each group has 6-7 topics (one has 26!) - each topic has 3 toggles - in-app, email and push That is 167 buttons to click! 😮 You can't turn off a group as a whole. And the gift keeps on giving - every time I turn them off, new topics pop up 🤦‍♀️.

Claude being honest about what it doesn't know. 👏 Thank you! > On the SHA pinning question — yes, I'm aware of the general trend of npm supply chain attacks. I don't have details on a specific recent axios incident since my knowledge cuts off May 2025, so I'd be making things up if I described it.

TIL You can download your LinkedIn data as CSV from www.linkedin.com/mypreference... I want to write a script to parse the csv, filter the connections in my network who I don't know (may have connected long ago). Then write a browser ext or Playwright script to unfollow/disconnect them on LinkedIn.

Very true! I see this first-hand in coding.

TIL you can nest markdown code blocks by using more backticks in the outer fence than inner one. I often want to document shell commands with their output - commands in ```sh, but output as plain text. I used to prefix each line with #. But now I use 4 backticks for outer block and 3 for the inner.

TIL: Create custom agents in Claude Code using /agents command to handle smaller tasks with specific tools, and can even run on smaller models. 🔥 Invoke in session: - @agent <name> - guaranteed to run - natural language - Claude decides Or at launch shell command - claude --agent <name>

Had a lot of fun at the OpenClaw hackathon at SF yesterday with @georgeck.me. Met a good mix of people - more senior devs than usual. We loved that the focus was to build something real. It nudged us to install OpenClaw on our laptop in a fully locked down docker container. luma.com/openclaw-hac...

Result from two harness/model side-by-side: Prompt: write a tetris game in typescript with beautiful UI Left: Claude Code + Sonnet 4.5 Right: Open Code + GLM-5 Harness matters. Model matters. What would it be with Claude Code + GLM-5?

There is something beautiful about sitting with Claude Code and working on a problem together. ❤️ Every time I talk to it, I learn something new. No judgment, we just brainstorm, try things, test and deploy! #joy-of-building

The line goes up and up 📈 2023: left corporate, went self-employed ❤️ 2025: started building personal projects 🚀 2026: discovered agentic coding 🔥 Three life pivots, three peaks. areyougoingexponential.rhys.dev/annjose

Levels of Agentic Engineering View Article | Join the HN Conversation Summary of HN discussion 🧵👇

TIL💎: VS Code has two super useful views in the left sidebar: 1) NPM SCRIPTS: shows all scripts across all sub-packages as one list. You can run / debug any script right there!🔥 2) TIMELINE: shows local history (undo, redo, saves) & git commits.💯 Right-click the EXPLORER header to turn them ON.

The @figma.com MCP server now supports a full roundtrip workflow from VS Code! Bring Figma designs into code, then capture your rendered UI back to the Figma canvas as editable frames. Join us for a special livestream tomorrow at 9:30 am PT to see it live in action: youtube.com/live/O1i-Te7...

Published a new skill: prompt-review It audits any prompt against Claude's official best practices - clarity, context, examples, structure. Outputs a scored review and a revised prompt. 🔥 It's part of my collection of agentic coding tools. Open source. Ready to use now! github.com/annjose/agen...

We often hear "AI is like a junior / intern" as if they're just eager but dumb. It always felt off coz juniors are usually smarter than we credit. I like this framing in Claude docs: > Think of Claude as a brilliant but new employee who lacks context on your norms and workflows The model context 💯

Played around with the new Nano Banana 2 and updated my avatar with an image it created. I selected a pencil sketch theme, uploaded my color profile photo, and it generated this version. Pretty impressive. blog.google/innovation-a...

Whenever I go to a web conference I like to see who begins eating their complimentary lunch first. I call this time to first bite.

Ideas were cheap. Not anymore.🔥 Coding agents have made it ridiculously easy to start building on any idea, reduced the friction to write code, super easy update it. You can bring ideas to life easier and faster than ever before.💪 So the most valuable asset is the idea. What a time to be alive!

A prompt to get an overview of a codebase makes the agent list the files they actually read vs. what was in context. In my exploration so far, this is the version works that gives meaningful info without burning too many tokens. gist.github.com/annjose/bc26... pi and Claude output below:

things you can do while waiting for the agent to implement what you asked for: 1. stretch back and legs 2. 10 squats, 10 push-ups, 10 lunges 3. drink water 4. walk around 5. read 6. maybe color, knit?? Now I can do what I have been trying (and failing) to do - take breaks between coding sessions!

Current coding workflow with 4 terminal windows: #1: pi coding agent with Codex for most coding tasks #2: Claude Code with Opus for big features that need detailed plan #3: OpenCode with Kimi 2.5 Free - ask trivial questions, brainstorm small ideas #4: Runs commands build, deploy So far, so good!🚀

I have been revamping many of my projects with agentic coding. I kept using the same prompts to scan repos, prep them and generate docs for agents. So I packaged that workflow into a new skill: agentify-repo You can run it from any repo. Added a few plans and specs too! github.com/annjose/agen...

I love this! data nerds, pls enjoy

Genuinely very impressed by the SVG of a pelican riding a bicycle I just got out of Google's new Gemini 3 Deep Think model simonwillison.net/2026/Feb/12/...

I had the hypothesis that harness is (as or) more important than the model, but couldn't prove it. Here is the proof! An excellent article. Must read. > The model is the moat. The harness is the bridge. Burning bridges just means fewer people bother to cross. blog.can.ac/2026/02/12/t...

We are still there in the Show HN main page - after 9 hrs of posting! 🔥 cc @georgeck.me ! news.ycombinator.com/show