Anjana Vakil

Yesterday, I put together a new community research project to study how the different agent platforms implement support for Agent Skills. The questions have direct implications for how we should author and publish skills intended to be portable across platforms. If you're a company publishing agent skills, you should be *very* interested in the results, because the implementation differences could be a source of bugs for your customers. You may not have realized you were signing on for it by putting out a few Markdown files that look like docs, but you have implicitly agreed to support the 25+ agent platform "runtimes" for your skills. That means we need to understand how those platforms diverge. But when I sat down in my kitchen yesterday with my morning coffee, I didn't intend to start a new research project. I was just checking a comment thread where a developer had requested a new feature in the `skill-validator` tool I maintain. How did that feature request turn into a GitHub project where I've created infrastructure to ask community members help me test real skill behavior across 25+ platforms? It all started with the greatest developer nerd-snipe of all time... someone on the internet was wrong. Link to the full writeup in the comments.

26

6 Comments

I added a parallel interpreter to the Kestrel library (https://lnkd.in/gXwMTfTj), allowing subqueries to run with parallelism, while maintaining the same high-level API for users. This is using Unison's green threads, async programming support, and abilities, which let you write straight-line code and have it turned into a pretty intricate program that does parallel processing.

16

I turned 3 projects worth of AI agent lessons into a single npx command. My last post about building AI development teams got a lot of attention. The number one question: "How do I actually set this up?" The honest answer was painful — it took half a day just to configure Project C. Eight agent definitions, skill files, file ownership matrices, settings... all by hand. And if Claude Code's spec changed? Your carefully crafted prompts were already outdated. So I built create-agent-system. npx create-agent-system One command. Pick a preset. Get a working agent team. What it does: - 3 presets (solo-dev / small-team / full-team) — opinionated defaults, not blank canvases - 8 agent types with role isolation and file ownership - 8 skill packages (scoring, TDD, visual QA, code review, ADR writing...) - Intersection skill computation — agents only get skills relevant to their preset - /sync-spec skill that validates your setup against the latest official docs The key design decision: scaffold once, stay up to date forever. Most scaffolders generate boilerplate and walk away. create-agent-system bundles a sync-spec skill that checks your configuration against  the latest Claude Code documentation via Context7 MCP. Your agents evolve with the platform. The technical highlight I'm most proud of: intersection skill computation. Each agent has default skills. Each preset activates certain skills. The agent gets the intersection — not the union. Why? Because irrelevant skills pollute the agent's context. I learned this the hard way when a Backend agent started demanding screenshots for visual QA checks. 5 things I learned building this: 1. Good defaults are the best documentation 2. Presets are opinions — opinionless tools are unusable 3. Scaffolding without sync is half the job 4. More agents = more important to filter skills precisely 5. There's a right level of abstraction — 100-line Handlebars templates are not it It's open source: https://lnkd.in/gmHwBYfA Full writeup with design philosophy, preset comparison, and code examples: https://lnkd.in/g62cPXCX #ClaudeCode #AI #OpenSource #DeveloperExperience #AIEngineering

4

I have been exploring TOON lately, and it feels like a smart upgrade from JSON for LLM workflows. "TOON" keeps the structure we already know, but the notation is lighter. That simple change makes prompts smaller, cleaner, and easier for models to understand. Fewer brackets, fewer symbols, and fewer tokens. That means lower cost and faster responses. What I really like is how naturally it fits LLM thinking. TOON is readable for humans and predictable for models. If you work with agents, prompts, or any system where you pass structured data, TOON can save you real time and real money. https://lnkd.in/dP68cSvy Excited to see how far this format goes next. #TOON #AItools #LLM #PromptEngineering #StructuredData #DevTools #AIAgents #Productivity

6

LLMs are not anthropomorphic, they are prediction machines. I came across yet another article evaluating how helpful LLMs are as a couples counselor, this time at NPR (https://lnkd.in/gu6i2tHm). Emma Bowman, the author of the article, describes a situation with her partner involving messy emotions and how she reached out for help to ChatGPT. The author talks about how the AI agrees with her, is sycophantic and assigns other human traits to and judgements on the model. She even reached out to Myra Cheng for a statement - the reasearcher advises that the reason for the odd behavior of the model is the "huge American and white and male bias" of the internet, that is the data it was trained on. Without going into the complicated topic of biases on or off the internet, it is helpful to remember that LLMs are first and foremost probablistic prediction machines. If you use an LLM to write code, instead of answering questions which really a licensed therapist should answer, the machine will still be strongly inclined to follow the approach you began. Please spend some time watching Andrej Karpathy's excellent series of youtube videos https://lnkd.in/gfdus-Ut He will show you how to start from a blank sheet of paper and create a piece of software which writes poems in the style of Shakespear. Huggingface also has a number of intro courses on LLMs https://lnkd.in/ge5FxB4a There you will learn that the way an LLM responds to your question is by (jazz hands) predicting the most probable words and sentences which follow the words and sentences of your prompt. Think of it as giving the LLM "get your ducks in a" or maybe you give it "to be a a fly on the" and the LLM knowing to respond with row or wall. From this vantage point, the "sycophancy" of your AI counsellor will be easier to explain than by trying to go into the rabbit hole of biases. If you live in the US, your therapist speaks English, was educated within a US licensing program, and reads news from US outlets. If anything, I would expect the therapist to be more biased than the LLM, simply because a therapist can only consume so much information while the LLM's only limit is the size of the data center it resides in. I invite you to be sceptical of opinions or takes on AI which ignore the underpinning principles of how LLMs work.

14