I'm teaching three free lessons on Maven, starting September 29. Each is an hour long, with a live walkthrough and a recording available afterwards.

Never thought I'd be getting away from the haze and pollution in Singapore (because of the Indonesian fires) by visiting - wait for it - Delhi? Clear blue skies, here I come! 😂

I love building with AI. But what I am happy to delegate may be the very thing someone else loves doing. For some people, the act of writing code, composing music or making an image is part of what makes the work worth doing. A finished result is only part of the story. This edition of Super IC asks what happens when AI separates a practice people love from the livelihood that lets them spend their days doing it. What part of your work would you miss if you no longer did it yourself? superic.ai/p/what-if-the-wor…

Some of the most useful instructions we give AI arrive as corrections. “This was a proposal, not a decision.” “That number needs a different denominator.” “There isn’t enough evidence to make this recommendation.” Those corrections contain our judgement. Yet we often repeat them in the next conversation instead of turning them into a method the agent can reuse. This edition of Super IC explores how agent skills can capture that expertise - and how to check whether they actually improve the work. superic.ai/p/your-expertise-…

AI can make each of us faster without making the team faster. If I pass an unchecked AI-generated document or piece of code to a colleague, my time saving becomes their extra work. They have to find the gaps and make the decisions I skipped. In this edition of Super IC, I explore the working agreements that turn individual speed into shared progress: visible decisions, clear ownership and work others can trust. superic.ai/p/individual-spee…

I cut back from running five or six AI agent sessions at once to two. I was getting more output, but the quality of my decisions was slipping. Building with AI puts you in an unusual position. You need the sustained concentration of a maker while handling a manager’s stream of decisions. What should the agent build? Is this approach sound? What did it miss? Can I trust this enough to move on? As the work speeds up, those decisions arrive closer together. Fatigue leads to weaker choices, which create rework and yet more decisions. In this edition of Super IC, I explore what this means for how we build - and how teams organize the working day. We need to account for the attention and judgement that all this output requires. superic.ai/p/the-decision-fa… (Oh - and I started a new newsletter. Check it out!)

11-year-old this morning in a super serious tone - "you can actually tell a lot about a book by its cover".

Every agent I use now (Claude, ChatGPT, Grok) can reason well. None of them do a task the same way twice. I noticed this doing something boring: asking an agent to review a contract the way I do it, with my checklist, my exceptions. Tuesday's output didn't match Thursday's. Not because the model got worse. Nothing ever told it my procedure, and it forgets my last correction the moment the session ends. That's not a model problem. It's a missing-skill problem. A skill is a small file that teaches an agent your procedure once. Claude, ChatGPT, and Gemini all read the same format, and it stays out of the way until the task actually matches. There are two ways to get one. Adapt what already exists: skill repositories now hold strong, well-tested skills for real work like code review, data cleaning, and documentation, so find one close to your workflow and adjust it. Or build one from your own expertise, the stuff no registry has because no one else does your job your way. That second kind takes longer and is worth more, but neither is the lesser option. Building it well has an order that's easy to skip: interview yourself for the gotchas nobody wrote down, write the description before anything else since that's the only part the agent always sees, write the body next with nothing generic in it, then add two or three real evals so you're not just guessing it works. Most skills fail because someone skipped that last step. I'm running a live workshop on this on September 17, 10:30am to 2pm EDT. We build a real skill from a task you actually do, test it, and ship it in one sitting. maven.com/anshumani/agent-sk…

This is the full primer behind my Agent Skills masterclass, a practical, no-fluff reference for building, evaluating, and shipping skills across Claude Code, ChatGPT and Gemini. You'll get the open Agent Skills standard and SKILL [.] md file format; a clear mental model for skills vs MCP vs subagents vs plugins; platform-by-platform breakdowns for Claude Code, ChatGPT and Gemini, including their extended frontmatter and quirks; a full workflow for designing a new skill from scratch, sourcing real expertise, and avoiding the generic-instructions trap; techniques for writing descriptions that trigger reliably instead of silently failing; an eval-driven development framework for testing and scoring skills before you ship them. 23 pages, written for anyone (technical or non-technical) building or maintaining agent skills. Share with friends, colleagues and other builders. Agent Skills are for everyone. maven.com/anshumani/o/fa0c8b