Claude Code may feel worse with Opus 5 because the skills, hooks, and `CLAUDE.md` instructions built for older models can now get in its way. I break down Boris Cherny's advice, show what happened when I tested Claude with and without my AI OS context, and share the three changes I'm making: simplify old instructions, prompt at a higher level, and give the agent a way to verify its own work. The goal is not to delete everything. It is to keep the context Claude needs and stop micromanaging how it gets there.

Hi everybody! I created an ecommerce competitor website scraper tool. Basically, what it does is scrape a list of competitor URLs I provide, pulling data across these areas: - Offers and pricing - Site structure - Product Page CRO Signals - Messaging - Brand & Visual Identity I started in plan mode, and while iterating I realised the CRO part would be much easier to follow with image evidence. So Claude created a new folder with evidence for each competitor. On top of that, Claude built an interactive/visual HTML file covering the key aspects, along with the full visual identity (colors, typography, logos,...). Two learnings from this challenge: 1. Plan mode is really important before executing any task. For example, before activating plan mode I asked for the Firecrawl cheatsheet .md file, Claude not just created the cheatsheet but also created a claude.md file in the same action without me asking for that. 2. As Nate says in the video, test with just one competitor first and iterate until it's polished. I started with 4 urls to compare outputs and it was tedious (it took a lot of time). Really enjoying this 7-day challenge!! #AISChallenge

Grok Bot makes it surprisingly easy to spin up a team of specialized AI agents that stay synced across your computer and phone. In this video, I walk through agent computers, teachable skills, scheduled routines, Slack triggers, agents messaging each other, and a live multi-agent website build. I also break down where GrokBot fits next to Claude Code, Codex, and Hermes, plus the limitations you should know before building your whole workflow around it.

Over the past few days, I’ve been working on a solution to efficiently summarize and search my YouTube subscriptions. After building my AIOS (AI Operating System) and "Brain Matter" (my personal knowledge base) over the weekend doing Nate's AI OS lesson, I’ve developed a workflow last night that works exceptionally well. Attached below is an infographic detailing the skill's capabilities, a progress in Obsidian as I further refined the skill, along with screenshots of the Markdown files it generates. A key feature is the prompt system—it allows you to choose between a fresh data fetch, grabbing the latest videos, or performing a local search, which saves significantly on token usage. I’ve also included "in progress" shots of my Brain Matter in Obsidian. The initial ingest processed 11,500 videos from 41 channels from over the past year, in just 10 minutes. Future fetches will only save new videos summaries into the raw folder then ingest them automatically. I have included a summary overview of the entire process so you can create your own.

Day 7 done. I built my executive assistant for day 7. This executive assistant didnt do good on first try, one thing I worked on to make sure that it works well is letting it learn my content style. I eventually gave it an example of what I'm expecting or how I want the script to be constructed for my short-form videos and long-form videos, and how it can turn my long-form video script to 3-5 short scripts. I tested it out and it really works great for me. I'm excited about this milestone. 🤩