💥NEW: CODEX POWER USER GUIDE 3 rules that Codex power users follow to build better and get more done. Rule 1: The Verify Rule - Give Codex a way to verify its own work using browser use, computer use, and tests (and TDD). - Frame your prompts as a goal (outcome + constraints + when to finish) instead of a task. - Require Codex to gather evidence (test outputs, screenshots, exit codes) to prove success. Rule 2: The Context Rule - Onboard Codex like you would a new employee or teammate. - Use Plugins to connect Codex to the tools and systems you use for development and knowledge work. - Create your own Agent Skills to encode business processes and workflows + make them legible to Codex. - Login to SaaS tools via Codex's in-app browser. - Combine plugins, browser use and skills with Scheduled Tasks to put meaningful work on autopilot. Rule 3: The Orchestrator Rule - Let Codex prompt and manage itself (Codex is better at prompting Codex than me or you). - Give Codex an outcome and ask it to spin up one or more threads to achieve it. - Use ChatGPT Voice Mode to talk to Codex like a chief of staff, have it spin up new threads to do work and report back with results. - Ask Codex to use subagents for planning, implementation and review. Any Codex tips or best practices I missed that you'd add? Watch my full full video guide (with examples and demos):

The X Subscription might be the most value-dense subscription of 2026. Insane to get access to Cursor, GrokBot, X all in just one subscription (very reasonably priced I might add).

Every book should come with a set of Skills to give your agent to help you better absorb and apply the content.

A new Skill I'm enjoying using is /blast-radius. I made /blast-radius to understand the magnitude of a change before I make it. It maps what a change touches and what it could break by exploring the project and codebase. /blast-radius was inspired by my experience as a PM. A common ask was sizing effort, so knowing the blast-radius of a feature was helpful, especially if it touched multiple teams or products. /blast-radius reads the code and maps out what it touches, how big of a lift it will be, and what the dependencies are:

The GMA System for building with coding agents (like Claude Code, Codex or GrokBot) is super flexible. You don't have to follow every single step for every project or feature you're building. Instead, just pick and choose the phases and skills that match the size of the task and the problem you're dealing with. Here's my rough breakdown: - Go straight to building: Sometimes I skip straight to the build. I just tell the Agent what I want and build it. No need to overcomplicate things. - Small task: I often go straight from a brainstorm to building. Everything I need is actually in the brainstorm and the models are smart enough to fill in the rest. - A feature: For larger features, I go through the whole process: /brainstorm, /create-spec for a feature spec, /create-issues and potentially run /blast-radius, then build away. - Big project: When I'm starting a new big project, I brainstorm it, create a project spec, break that down into multiple issues, then build each issue. I illustrate this visually in this diagram:

/create-issues is a simple Skill that makes my building process much more structured and scalable. /create-issues takes in a project spec or other input and creates issues in GitHub or Linear. I mainly use it in 3 ways: (1) From a project spec: you get a bunch of issues, plus the dependencies between them, so you know which to build first. (2) From a feature spec: you get one issue you can build. (3) From a chat message: you don't even need a spec. Dump a bunch of screenshots and say, "Here are the bugs in the app. Please create issues to capture all of these bugs, and let's tackle them in a next session." This makes it easy to create a bunch of issues at once. One of the most important parts of the issue template is the acceptance and verification criteria, which tell your Agent how to verify that the feature actually works as intended.

"I could have learned Claude Code on my own, but it would have taken much longer to reach the same level of confidence." - Mario, ML Ops Engineer Claude Code 2 Day Launchpad starts this Friday. It's a live building experience for people who want a better system for their work with Claude Code. How it works: - You'll build a project of your choice with Claude Code through a series of mini-projects. (It can be a new app or a workflow for your job / business) - You'll learn a repeatable system for building and Claude Code fundamentals that will make every future project easier. - You'll get live help from me, plus a supportive group of peers also building live. - You'll walk away with real confidence to use Claude Code in your work, whether its building products with AI or AI-enabled workflows. Learn more and grab your spot here: claudecode101.ai

Models I use the most: 1. Opus 5.5 (on medium / extra high effort) 2. GPT-6 Astra (medium / ultra) 3. Everything else

"Plan mode was useful, and is no longer useful." - Boris Cherny (@bcherny), creator of Claude Code Plan Mode is a mode in Claude Code that doesn't write code. It takes in your prompt and makes a plan for you to review and agree on before it builds anything. I used to use Plan Mode as my default way of working. But recently I found that I no longer use it at all. My reason is that the models have gotten so much better that I don't need Plan mode anymore. This puts into words something I've been feeling: models like Fable and Opus, especially Opus 5.5 in Claude Code, have gotten to the point where you don't need Plan Mode anymore. Instead of Plan mode, I've adopted the @steipete method of "Just Talk To Your Agent". I have a set of Skills which facilitate a conversation and help get clear on the goals, constraints, and verification criteria before we kick off the build