This was such a challenge. I had a idea for a hero landing page design but it was not coming together as I hoped. It got to the point where i was using GEMINI, CLAUDE and CODEX to "just get something close it". Alot of back and forth prompting, folder structure fell apart the first time. I just had to keep engineering it. Tried it forward and reverse engineering it. I was patient but very persistent. I COULD HAVE BEEEN BETTER ORGANIZING AND STRUCTURING MY FOLDERS. It seemed like there was no hope. It seemed no matter how good i got them, the design was just odd, not as direct and very confusing. WHAT DID THIS DO....? ACTUALLY, it inspired me to create my own tools and workflow systems so my designs can manifest exactly how i please and i can have more control of the details. Something ill be speaking and following @Ari Evergreen and @David Vogel about. And of course more videos from @Jake Van Clief. I went into gemini and created parts of the design and had it add controls for it so i can get it looking how i wanted, then took the code and added them to my files or my prompt. This seemed to really advance the design forward. i need to find a better way to update the files. This design was not easy! I want to thank @Shirsho Guha @Marcos Accioly and all others for the feedback on the my other post for my first website. You can see the old version and 1st post about it here: https://www.skool.com/quantum-quill-lyceum-1116/project-2-personal-website-sheesh?p=8319ce31 Definitely learned alot and look forward to adjusting for better workflow and designs. V5 with HERO DESIGN https://www.koachkev.io/

Inspired by @Curtis Hays, I finished an experiment that I had started a few weeks ago. I started my company as a local web design agency in 2014. I've pivoted a few times since, where our core offer is that of Fractional CMO services, but we still do some agency type work for existing clients, friends, and family. I was ready to throw in the towel on web builds just because they take so much time and I hadn't cracked the code on building sites with ongoing content. I finally got it this evening! We've gotten good at building unique sites using Claude with next.js, files live in github, website runs on vercel. This evening I was able to get a headless WordPress environment where I can add/edit/delete content, but the design still lives in Vercel! I realize this might seem like a small thing for some, but it's a ginormous win for the 8 SIGNAL team :D It's gonna save us sooooooo much time rebuilding from next.js to Elementor. Huge shoutout to Curtis. By sharing his win, he inspired me to dig further into this and figure it out! Here's the site that was previously just a "static" site on vercel, now all the content is coming from a wordpress install: https://century-rentals.vercel.app/ P.S. If you have a better way of managing websites that require a CMS and SEO/AEO, don't hold back. I won't be deflated. I'm always looking for ways to improve and do things better. P.P.S. Website's for a buddy, and still not finalized yet, so I'm taking some liberty to mess around with him about baseball because he knows less than I do, and I don't know much LOL!

Meta shipped a connector for Claude and ChatGPT on April 29 that runs write operations directly on your Facebook ad account, no API keys, no app review. Okay the title is half clickbait, but the substance behind it is the part nobody is covering correctly, so here goes. First, understand what Meta actually shipped. It's two surfaces. There's the Meta Ads MCP server, which is what you point Claude or ChatGPT at to control your ad account by talking to it. And there's a Meta Ads CLI, which is the same capability surfaced as a command line tool for engineering teams building internal pipelines. Both are in open beta, both are free, and both authenticate through a normal Meta login. That last part is the unlock honestly, because every prior attempt at agentic Meta ads required going through the Marketing API, which means dev credentials, app review, and business verification. The MCP server collapses that whole credential dance into clicking "log in with Meta." The capability itself is split into four buckets and they actually let you write, not just read. Reporting is the obvious one, pull performance, surface insights, the standard stuff. Campaign management is where it gets interesting because the agent can create and edit ads, ad sets, and full campaigns by natural language. Catalog management lets it build product catalogs and fix data feed problems. Signal diagnostics covers Conversions API health, learning phase status, and signal prioritization. Most third-party ad-MCP wrappers (AdKit shipped one for Google plus Meta on May 1, two days after Meta's own launch) are read-only, because writing to ad accounts through unofficial APIs is risky and Meta gates it hard. Meta's own connector doesn't have that limitation because Meta is the platform. Now the part where I'm not just selling it. It's open beta, so the rate limits and edge cases aren't fully documented. The bigger thing to know is the learning phase, which is roughly 50 optimization events that Meta uses to actually figure out who to show your ad to. If your agent edits a campaign mid-learning, the campaign resets and your performance tanks. Meta hasn't built guardrails for this yet, which means the agent has zero idea by default. You have to tell it. If you don't, you'll watch your CPA spike on the first day you let Claude touch your account and blame the connector when it was actually you. (fyi this is the same thing that happens to humans editing campaigns mid-learning, the AI just does it faster.)

This is a quick one. I took for granted that you guys already knew this. But maybe you don't? It's pretty great. In Claude Code / Cowork, have Claude spawn agents to critique your code, writing, anything. You can even define who they are. You will get a much better idea of what you've written from several points of view. I get WAY less sycophantic feedback from the agents. They will even find code problems and writing inconsistencies that default Claude wont. And they can get feisty! For instance, you can give your Clief Notes post to Claude and say, "Spawn 5 agents that represent the Clief Notes community. They should all have strong opinions. Have them critique this post." I gave everything above this line to Claude and then ran the agents, so they don't know THEY are the examples. haha ;) Here's what they said: --- Agent 1 -- The Builder: Solid concept, but this reads like a tweet stretched into a post. You're telling us that it works without showing us how it works or what it actually looks like. Drop in a real example -- show the prompt you used, show one of the agent responses, let people see the feisty feedback you're bragging about. "Spawn 5 agents" is doing a lot of heavy lifting with zero specifics on what makes a good critic prompt versus a bad one. The idea has legs, but right now it's a teaser for a post you haven't written yet. Agent 2 -- The Skeptic: You're telling me agents "get feisty" and "find code problems" -- show me. One before/after example would do more than five sentences of hype. Also, "spawn 5 agents" is doing a lot of heavy lifting here -- are you prompting each one with a distinct persona, or just asking Claude to roleplay five slightly different yes-men? Because without deliberate persona design, you're just getting the same model talking to itself in different fonts. The idea is legit -- adversarial feedback loops are genuinely useful -- but this post asks me to take your word for it instead of proving it. Agent 3 -- The Newcomer:

jcode boots Claude Opus in 14 ms, runs at 27.8 MB, and edits its own Rust source mid-session. Same model inside. I believe this is the start of a sub-genre nobody is naming yet. Quick context: The first era of LLMs was prompt engineering. Era two was context engineering. Era three is what we've all been arguing about all year, model picking, Opus 4.7 vs Gemini 3 Pro vs Kimi K2.6. Now, someone rebuilt the wrapper around Claude in Rust, and the gap on cold start, RAM, and per-session scaling is bigger than any model gap I've seen this year. What jcode actually is: - Coding agent harness, 94% Rust - Custom terminal called handterm, custom Rust mermaid renderer - Native logins for Claude, OpenAI, Gemini, Copilot, Azure, plus aggregator providers - MCP works out of the box. Falls back to your existing .claude/mcp.json so anything you've already wired up keeps running - Install via brew, curl, or cargo build. Single binary - 3.3k stars on GitHub, 57 releases, actively maintained The numbers (jcode vs Claude Code, from the README): - Cold start: 14 ms vs 590 ms to 3.4 seconds. 42 to 245 times faster - Idle RAM with local embedding off: 27.8 MB - 10 sessions in parallel: 260 MB total vs 334 MB to 3.2 GB for Claude Code - Per added session: 9.9 MB vs 76 to 318 MB - Custom mermaid renderer the author claims is 1800x faster than browser-based versions That last one is the kind of detail that tells you what they're really doing. Someone is going through every layer of the Claude Code experience and rebuilding it natively, and the gains compound. What's actually different at the harness layer: - Self-dev mode. Agents inside jcode can edit the harness's own Rust source, run cargo build, hot-reload the binary across active sessions without dropping you. The wrapper is recursively modifiable from inside the agent loop. - Memory as semantic vectors per turn. Recall is automatic via cosine similarity. Not "remember to update CLAUDE.md". There's an ambient mode that consolidates memory and resolves conflicts in the background while you work.