Lawrence Wu

The AI Engineer World's Fair by @aiDotEngineer is a great conference I attended a couple years ago. I couldn't make it this year but wanted to see which talks were good. They've started posting the talks on YouTube so I had Codex/GPT5.5 build me a ranked list of most popular talks by views per day. I have to remind myself that these tasks to scrape some data to help some decision is trivial for coding agents to do now. Here's the post showing how this was done: lawrencewu.net/posts/2026-07-… These are the top 5 talks so far (out of 77 posted): 1. What do we build now? - Theo Browne, @t3dotgg Views: 25,835 Views/day: 25,835.0 youtube.com/watch?v=xUnRQ9… 2. Field Guide to Fable - Thariq Shihipar @trq212, Anthropic Views: 47,730 Views/day: 15,910.0 youtube.com/watch?v=9fubhl… 3. Building Great Agent Skills: The Missing Manual - Matt Pocock @mattpocockuk Views: 93,712 Views/day: 9,371.2 youtube.com/watch?v=UNzCG3… 4. The Future Is Domain-Specific Agents - Justin Schroeder, StandardAgents Views: 28,792 Views/day: 2,879.2 youtube.com/watch?v=spNAUE… 5. Building an ACP-Compatible Agent Live - Bennet Fenner, Zed Views: 2,533 Views/day: 2,533.0 youtube.com/watch?v=HsxQIC…

Claude Opus 4.8 was released yesterday. The comparisons to Opus 4.7 and GPT 5.5 were quite strong. I was surprised Anthropic didn’t put Mythos Preview as a model they are comparing against. I added Mythos Preview and Opus 4.6 to the graphic. Mythos Preview is still beating Opus 4.8 by a healthy margin on these benchmarks. There are rumors a Mythos-class model will be released in a few weeks. Gist of the SVG: gist.github.com/lawwu/876d172d… Opus 4.8: anthropic.com/news/claude-op… Mythos Preview System Card: www-cdn.anthropic.com/8b8380204f7467…

@BrentBeshore Thanks for sharing! Very moving. What a well lived life that God used to shape his daughter into the woman she is today.

What a beautiful reflection by Alex Sasse, @BenSasse's eldest daughter. Full of deep wisdom. Can't imagine a better use of 7 minutes. Link below. "Every single day of my childhood, my parents asked the same question at dinner. Not 'What did you learn?' but 'Who did you serve?'"

I did a 2+ hour hands-on AI tutorial for some friends — no coding background required. If you've used ChatGPT, you'll understand this. And hopefully you'll see why coding agents are so much more useful to do real work. We covered: - The difference between ChatGPT and coding agents like Claude Code - What "bash" is and why it gives AI superpowers - Real demos: Excel data, dashboards, video editing, ML pipelines - Token costs, model choices, and how to save money - How to run AI models locally on your own Mac The goal was to show normal people (not developers) what's actually possible with these tools today — using plain English, no code needed. Video: youtube.com/watch?v=zh5twq… Resources: github.com/lawwu/ai-tutor… #AI #ClaudeCode #ClaudeCowork #ArtificialIntelligence #Tutorial #Video

So timely, I was talking with one of my colleagues today about this exact topic. We were discussing how using Claude Code now is pretty exhausting and how work is different with Claude Code. You are processing so much more information. I'd estimate it's 3x more mentally taxing than working without Claude Code. I totally agree with @simonw one can grow in this skill of working with coding agents but there are human limits to what we can do with a finite amount of time and energy. It'll be interesting how different organizations set expectations on their engineers? And engineers at different levels. What should an entry-level software engineer be expected to do?

"Using coding agents well is taking every inch of my 25 years of experience as a software engineer, and it is mentally exhausting. I can fire up four agents in parallel and have them work on four different problems, and by 11am I am wiped out for the day. There is a limit on human cognition. Even if you're not reviewing everything they're doing, how much you can hold in your head at one time. There's a sort of personal skill that we have to learn, which is finding our new limits. What is a responsible way for us to not burn out, and for us to use the time that we have?" @simonw

@mattpocockuk Used your /grill-me skill for the first time yesterday, I agree it is a lot better than Claude's default plan mode.

There was some confusion around this, so let me clarify. - I don't use plan mode - I still plan like crazy, using my skills /grill-me, /write-a-prd, then /prd-to-issues - Bad plans = Bad outputs

I did a live AI tutorial for some friends a few weeks ago. The recording wasn't really good because my OBS settings were right. I created a better version of a tutorial called a Deep Dive into Coding Agents like Claude Code. 00:00 Introduction 08:01 Agent Evolution (2022–2025) 14:25 Going deeper and broader 22:39 Live demo 27:08 Cowork vs. Claude Code side-by-side 36:55 Core concepts 48:40 Skills 50:03 /skill-creator with built-in evals 01:01:45 Real-world examples youtube.com/watch?v=qWjw3c… #claudecode #codingagents #knowledgework

This week's newsletter: lawrencewu.substack.com/p/three-years-…

It's hard to believe it's been 3 years since my first blog post — which was a simple list of Mac apps I thought were useful. Since then I've become convinced that writing is one of the most underrated skills in the agentic knowledge work era. This week's newsletter covers: - the top 10 human skills for agentic knowledge work - why Karpathy's autoresearch is the first true killer app for coding agents - and a few reflections on 3 years of writing in public Link in the comments 👇

Karpathy's autoresearch is the first killer app I've seen for Claude Code. It runs 50-100 training experiments in a short loop, letting an agent reason about what to try, write the code, run it, and iterate. The kind of thing that used to take me weeks of manual parameter tuning. I've successfully used it to improve models I've worked on by 15-20%. The models weren't bad before — I just couldn't justify the cognitive overhead of trying every variation. autoresearch removes that bottleneck entirely. And the usage is almost embarrassingly simple: Prompt Claude Code to "use karpathy's autoresearch to improve this model" That's it. This is why I think coding agents are a genuine step change. autoresearch needs both a model's reasoning and an agentic harness to actually run code. Neither alone gets you there. Seeing them come together like this is exactly what I hoped agents would eventually unlock. Years ago AutoML was supposed to do this. autoresearch is the first thing I've seen that actually delivers. → Post: lawrencewu.net/posts/2026-03-… → github.com/karpathy/autor… → github.com/alvinunreal/aw…

Man supply chain attacks are so horrible. I'm guessing these attacks will become more common and more creative as LLMs and AI Agents allow attackers to scale their activities so widely and creatively.

In the era of agentic knowledge work, AI isn't replacing human skills — it's amplifying them. Here are the top 10 human skills I think are currently most valuable: 1. Mastery of AI tools 2. Reading 3. Writing 4. Understanding what AI can do 5. Communication 6. Curiosity & eagerness to learn 7. Humility 8. Coding / programming 9. Domain expertise 10. Optimization mindset See more details in this post: lawrencewu.net/posts/2026-03-… Curious what you think are the most valuable human skills now?

Everyone is wondering how AI is going to impact jobs. Karpathy had a tweet I wholeheartedly agree with: "The art of the process is spending 80% of the time getting work done in the setup you’re comfortable with and that actually works, and 20% exploration of what might be the next step up even if it doesn’t work yet." I do try to ask myself very frequently: “What is the agent-first way of doing this?” I’ve had to rebuild so many different habits I’ve built up over 15 years of working. It's honestly been disorienting at work. I’ve been telling my friends and those close me that I think knowledge work is changing fast. So much digital work or work that can be done on a computer is going to be done by AI Agents or more specifically those people who have developed expertise in using AI Agents. And in 1-2 years, if you are not keeping learning how to use AI, specifically agentic AI, you will not be as employable in these white-collar jobs where you’re primarily using a computer. It takes time to develop skills in working with AI in your field. I encourage everyone to start now. Invest 10-20% of your time trying to reproduce your current work or workflows using AI. This is an excerpt from my weekly-ish newsletter here: lawrencewu.substack.com/p/agentic-know…

I put together a short video showing how to breakdown a whole brisket from Costco. Processing meat this way saves you about $220-260 per brisket if you compare store-bought sliced prime brisket ($15-18/lb) vs. slicing it yourself ($4-$6/lb + some effort). youtube.com/watch?v=Uk_6nq…

I’ve been spending time exploring agent skills and skills marketplaces lately. There are already some great examples out there from Anthropic, OpenAI, Sentry, Compound Engineering, and skills.sh. It's now possible to easily setup a skills marketplace for your own organization. That way your employees have a central place to contribute new skills and install and use existing skills. To make that easier, I put together a template for creating an Agent Skills Marketplace: github.com/lawwu/skills-m…. It's a Cookiecutter template that packages together some of the best patterns I’ve seen across existing marketplaces, including: - symlink conventions like CLAUDE.md -> AGENTS.md - shared skills directories - GitHub Actions for Claude Review, Skill Review, and frontmatter validation I’m also using it in practice here: github.com/lawwu/skills One thing I especially like is that you can bootstrap the whole setup through a skill itself, which feels very fitting: /create-skills-marketplace It automates much of the setup that would otherwise be manual. Enjoy!

I believe we’re entering an era of Agentic Knowledge Work. In this post, I share some observations on how knowledge work is changing because of agents like Claude Code that are surprisingly generalizable. I'm also seeing glimpses of agentic knowledge work happening in my field of data science and machine learning. lawwu.github.io/posts/2026-03-…

I just wrapped up sharing 15 days of Claude Code tips. I wanted to somehow keep this going so I’ve started a Claude Code Field Guide where I’ll document: - Claude Code best practices (at least for my use cases) - Skills, Plugins I like - Other things I’m learning about work and myself as I use these agentic coding tools Because the agent harness is changing so frequently, I’ll also document: - Things I’m currently trying - Things I’d like to try - Things I’ve stopped doing github.com/lawwu/claude-c…

"We build our entire harness around prompt caching" - Thariq (Claude Code dev) I've really liked Chip Huyen's sniffly project to visualize Claude Code stats. One of the graphs shows you the daily cost of your tokens by input tokens, output tokens and cache operations. I always thought this graph was wrong because it looked like 99% of the cost was cache operations. But I realized it wasn't wrong. I underestimated the importance of prompt caching. Cached tokens are 10% the cost of regular tokens (!). This means Claude Code structures static content first and dynamic content later in the prompt: 1. System Prompt & Tools (global cache) 2. CLAUDE.md (cached within a porject) 3. Session context (cached within a session) 4. Conversation messages I didn't appreciate how much engineering and effort goes into making the user experience in Claude Code what it is AND keeping the costs maintainable. If you are building your own AI systems on top of LLM APIs that support prompt caching, you need to design your system with prompt caching at the forefront of your mind. Looking at the Claude Code CHANGELOG, you can see there are dedicated releases for fixing things that led to reduced cache hit rates, e.g. 2.1.62. Thariq has a good post on the importance of prompt caching that was so illuminating: x.com/trq212/status/…

Don't allow the agent to outpace your understanding of whatever you are building. The models can still make mistakes. You need to have some understanding of the system in order to validate what the agent is doing. Your understanding of the system is also critical to guide future changes and validate correctness Practical tips: - Use plan mode (Day 6) for complex tasks — review the plan before executing - Read the code/diffs, don't just approve them - Ask Claude to explain its changes: "why did you do it this way?" - Keep AGENTS.md updated so future sessions have context on architectural decisions - When Claude writes code you don't understand, stop and learn before moving on This is even more important when using multiple agents (Codex + Claude Code) — you're the one who has to maintain the codebase (at least for now?) (Day 14 Claude Code Daily Tips)

I enjoyed reading all of Simon Willison's new guide for Agentic Engineering Patterns. This one in particular resonates in "hoarding things you know how to do." I've always tried to write things down and document things mainly because I didn't want to forget but that becomes even more important to allow coding agents to use agentic search to find things. I've observed this practically as I've learned how to use Github Actions productively over the last 6 months, seen the value of them and see the combination of Github Actions + Coding agents being really powerful. For example, setting up new repos to have a claude-code review Action involves setting API keys and GPG keys for git signing. I realized all of this can be put behind a Skill and that behind a Claude Code plugin that makes it easy for developers to add this functionality to their repo. As a data scientist, I've always thought of myself as a composer of different tools and abilities. Those tools used to be data analysis and ML libraries like dplyr, pandas, scikit-learn and PyTorch. Now those tools are Claude Code, Codex, Github Actions and Agent skills.