Macey Baker

My main takeaway is that skills are software, and the same rules apply. The things that make a bad skill also make a bad software component, eg. being badly scoped. "Should we factor this out" has become "Should we make a skill for this". Same stuff, different form factor

@rseroter @_philschmid This is excellent. It's a bit buried but I especially love the point about skill retirement. Skill bloat is coming :) A couple others have mentioned it but we genuinely do quite good plug-and-play skill evals at Tessl. Check out this one for example tessl.io/registry/skill…

"You wouldn't ship code without tests, but why ship skills without evals? This is a practical guide to fixing that." philschmid.de/testing-skills < @_philschmid making us smarter on defining success for Skills and iterating on evaluations.

no man talks to the same llm twice, for he is not the same man and it is not the same llm

Did you know that you can call agent skills *from* other agent skills? 😵‍💫 I wrote about this & other fun things you can do with skills here: tessl.io/blog/getting-c…

@basilYam Word.

I wish there was a pithy phrase that meant "I hear that you think this. I neither agree nor disagree. This is your thought, and I really have nothing to do with it."

Ok but *what* did the lymph knowed

@nonils @tessl_io Great write up. I completely agree with your assessment. Admittedly I took the scenic route to get there over the past year, but yeah, the up-front design-heavy process didn't make sense pre-LLMs, and doesn't make sense post-LLMs either. We're embracing this at @tessl_io too!

@macebake I really enjoyed your article about capturing intent, as some observations really resonate such as info trapped in discussions. IMO the answer is to derive spec from discussions. Would love to see @tessl_io move in that direction? meaningfool.net/articles/spec-…

When exactly did we stop saying "screen name"

@destraynor I was just recently reminiscing about the spaces vs tabs era.

There was once a time when the biggest, riskiest, most divisive conversations on Twitter were things like: * Tables vs <divs> for layout (and spacer gifs) * Should designers code? * Skeuomorphic vs flat design And they were good days tbh

@icanvardar Actually working quite well these days for small tickets and automations!

guys btw what happened to devin? wasn’t it supposed to take our jobs

Amazing how much we take air travel for granted.

@aruiz_tagle @chaisoyum @kindabored @MickyAbir @_opencv_ Tbf though, models are great at optimizing for those success markers, eg “cheating” on tests. “Code that works” is a really hard thing to evaluate at scale

@chaisoyum @kindabored @MickyAbir @_opencv_ Using a positive feedback loop of successful code/failed generations is arguably a much better way of refining an LLM than scraping the internet but we shall see!

Please teach LLM-backed tools about themselves. This is bare minimum now. I should be able to debug your tool *with* your tool most of the time

There’s a pattern with LLM dev tools: First few tries? Magic. Then you hit the edges: the LLM guesses, hallucinates, ignores your tools, blows through tokens… and the magic fades fast. @macebake shares what she learned building an MCP server for a CLI tool. MCPs shine when the user already knows what’s possible. But for new users? The limits show up quick. Check out the article for pitfalls + practical tips 👇 ainativedev.co/6yl #AINativeDev #ClaudeCode #MCP

On priming, a quick and dirty experiment with a `help` tool:

I wrote about some lessons learned building an MCP server... I imagine this will feel silly, for one reason or another, in a year's time. ainativedev.io/news/teaching-…

Just found out the sun has no proper scientific name. Wtf man

@stewartsc I'm glad it's not just me 😅

@macebake Thanks for writing this. This has been our experience too and appreciate the tips here a lot of which we haven't tried.