Mike Buckbee

Hey friend. Can you spare a ⭐️ for a (hopefully) interesting open source project I'm working on? github.com/ExpeditedPro...

Don't use the Chrome DevTools MCP to check a version upgrade. Have the agent script headless Chrome. Our Tailwind v4 check: 92 screenshots (23 pages, light and dark, v3 vs v4) and a pixel diff of each, in minutes. The MCP would need close to 200 tool calls.

I had Claude audit my running processes, startup items, unused apps, etc. ....soooo much cruft cleaned out of my Mac, it's like I just downloaded a new stick of ram into it.

Each week I try out one or two new AI skills/tools, mostly focused on the coding and productivity side of things. What new should I try this week? What are you using?

Fired up a grokbot to monitor new CVEs and map if they affect any of my repos as there's such a storm of them now it's nuts.

This explains my month

Google Trends is notoriously bad and it's hard to split out some of these terms, but this chart I think pretty indicative of the current relative adoption rate of Grokbot, Hermes, OpenClaw and Muse.

Imagine you have millions of documents with mentions of people, but not proper unique identifiers. How do you sort out Jane Doe 123 from Jane Doe 234? You use this amazing piece of FOSS software: Splink moj-analytical-services.github.io/splink/index...

Disposable AI Agents are going to be a "thing". With the big PHP security vulnerability and disclosure timeline there was a ton of information (good and bad) about what to do for mitigation. A Grokbot for a single day just to monitor sites/socials for the latest, hugely helpful.

Installing Hermes today. Anyone with tips/suggestions?

My winning entry for Hackoberfest is up, check it out.

RTK seemed like a really good idea (a local dev proxy utility that used leaner commands to curb token usage), but in practice the minimal saving aren't worth the added complexity and occasional screwups.

So LLMS.TXT - Anyone using it to point AI agents to where your MCP/signup or how to get started is located? It's on my list to add it to HutchDB today and would be nice to hear any real world tips. Thanks!

Claude code is now delegating to sub-agents like crazy so I've add a new timer based command to my claude file to give me an update on what they're doing every 3 minutes so I don't come back to a seemingly "stuck" command.

It's hilarious that all my speech to text tools "hear" Grokbot as Crockpot.

PSA: If you're using a displayport adapter to add extra monitors to your Mac, make sure you've updated the drivers. I just went from v14 -> v17 and it feels like I bought a new Mac it's so much faster.

Every Monday I ask Claude to make a 1 page PDF of my AI usage, new skills, what was unused and what to try in a futile attempt to keep up with the AI dev tooling changes that are happening.

What's best practice for organizing agents (Grokbot, Hermes, Openclaw?) One per task or one per role?

Find out which members of Congress actually voted for the issues you care about at dollarvote.us It's a site I put together for Hacktoberfest where local devs came together for a day of hanging out and creating cool stuff. It uses Google's Gemma AI to analyze in detail each bill and voting pattern

Entering the brave new world of hardware programming (as soon as my NFC reader + cards get here).

What's the term for a "fake applicant" who is applying for a role? Stuff like: "The person we interviewed had stolen a person's resume and was pretending to be them", but I don't have a succint term to wrap up this concept. Maybe "applicant fraud"?

Claude, Codex, whatever, go ask it: "Am I actually using all these skills and plugins I have loaded?" and it will grep your sessions and check for you. Just axed like half my setup.

It works! 173s -> 65s

I've been messing with text to speech models as well and it's still kind of mindblowing that hosted ones are significantly faster than locals.

Been making good productivity strides with adding some new skills to my Claude Code setup, anybody have any good recommendations for what they use?

Really trying to use AI Coding tools more descriptively+spec oriented way this week and less of my super handholding and iterative way that I normally do. Using `/grill-me` skill to get a lot more of the domain knowledge and questions up front. Anyone else have any tips?

It doesn't matter if you use Claude, Codex or Grok. If you're not building in continual improvement from each loop through the code you're going to eventually drown yourself in bad code.

What's a better Text To Speech? Say, Kokoro or paying OpenAI for API calls? I'm hacking together my own voice assistant and built a comparison of 25 curated options that are all pretty good. Personally probably going with af_nicole for this project. voice-bakeoff.vercel.app

What are folks doing for in-repo documentation for AI building? I've bounced between Google's OKF, a docs subfolder, overloading AGENTS.MD Any Tips?

You can ignore the posts about Jev and whether LLM classifers are better if you remember one thing: always split your AI classification and your AI generation into separate tasks.

Stupid low hanging fruit. Sped up our test suite by 5x by just asking Claude 'hey can you make the test suite faster'

Claude, Codex, Grok, etc. are wildly inefficient running shell commands because they call the default verbose versions of them. Using RTK, I've been able to cut input tokens from them by _half_ - nevermind that this also speeds up everything as well. Recommended.

Yesterday 7 of 3,727 user-facing AI API calls failed (0.19%), all upstream model timeouts. Users saw "try again later." Our p99 hit 3.7s, then recovered. Is 0.19% acceptable, or would you add a fallback model?

Are y'all running code review + testing agents on your code before sending it to PR or waiting until after?

Does anyone have any good examples of AI-assisted (MCP) onboarding for SaaS? Both Stripe and Posthog have gone this route but I'm trying to collect more as I think it's fascinating.

o11y hacks: add “number of visits to your status page” as an alert

Ever suffered through the pain of asking someone to make a DNS change over email? Suffer no more! I've just added the ability to share DNS change requests on the DNSisBeautiful site. Bundles the request + a test if they set it right into one shareable URL.

Tell me what's bad/wrong here? I'm prototyping a new feature for DNSisBeautiful that lets folks make request links via simple parameters that they can then send to someone to get them to make a DNS change. Would this make sense if you got landed on this page?

Used Opus 5.5 for the past 24 hours across ~5 projects and it sure does feel like Fable smarts but much faster.

If you want to try the new Opus 5.5 with Claude Code make sure that you've installed it via the native installer and not via homebrew (which lags releases).

It's kind of crazy how much better Claude Code is than Codex at handling and working with multiple different iterations of MCP servers. It's not even the "future" it's literally just what I need to get through my working day.

Anyone have a really good way of using worktrees in dev? I've got my own per-project setup but it feels like there's got to be a better way.

My new sign off on drip emails is "p.s. if you reply to this email a human will answer it." Is that actually helpful information to people signing up?

If you find yourself with extra AI token quota before a reset, start grinding out dependency+security updates and improving test coverage of your apps.