In addition to leaking neuralese, models can have bothersome verbal or behavioral tics. I can imagine a future where users only interact with "chief of staff" agents that embody and enforce user preferences, delegating all work to other agents and acting as a filter.

The state of agentic coding right now:

The skills​.sh API is now generally available. Power your agents, applications and platforms with access to over 600,000 skills ↓ vercel.com/changelog/the-ski…

Grok now has skills

Some people argue that skills will become obsolete as models improve. That may be true to some extent, but models will never know exactly how you (or your company) want things done. Personalization is a fantastic (and often overlooked) use case for skills.

I’ve seen people on X dunking on folks like @garrytan @doodlestein and others for sharing SKILL dot md files they've built. They are dismissing these files as "just a markdown file.” I think this misses the point entirely and I'll try to address that here. Quick thread: A bad skill file is just text, sure. A good skill file is compressed expertise, packaged in a format an agent can actually use. The value is not just in the “markdown file.” The value is the interaction between: a huge neural network with latent capabilities a precise, reusable, agent-readable procedure that steers those capabilities toward a specific outcome That combination is the product. Saying “it’s just markdown” is like saying Hamlet is “just ink on paper,” or Einstein’s relativity paper was “just a text.” Technically true. Intellectually useless. The medium is simple. The content is what matters. And more importantly, the effect of that content on the reader is what matters. With humans, a book, a coach, a lecture, or painting can change how someone thinks and acts. With LLMs, text is also the control surface. These models were trained on text, reason through text, call tools through text, and follow procedures through text. So yes, the skill is “just text.” But it is text designed to be read by an enormous neural net. That matters. A good skill is agent-ergonomic. It does not merely say “do this better.” It encodes workflow, constraints, examples, edge cases, tool usage, failure modes, and success criteria in a way the agent can reliably execute. That is very different from a casual prompt. A prompt is often a one-off request. A skill can be reused, versioned, tested, improved, shared, and loaded at the exact moment an agent needs it. That turns “vibes-based prompting” into something closer to operational knowledge. Another way to think about it: We have built these massive models, but much of their power is latent. Different people can extract very different levels of performance from the same model. A good skill is a way to actualize a specific slice of that latent capability. A refactoring skill. A research skill. A legal review skill. A math explanation skill. A codebase-navigation skill. Each one can make the same model behave very differently. I think of Cus D’Amato and Mike Tyson. Tyson had enormous latent potential. But Cus gave him a system, a style, a discipline, a way to channel that potential. That’s what good skills are for agents. They are not magic. They are not all equally valuable. Many will be mediocre or useless. But dismissing them right off the batt because they are “just markdown” shows a misunderstanding of what LLMs are. Text is how we trained these systems. (for the most part) Text is how we steer them. Text is how we unlock parts of what they can do. The question is not whether a skill file is “just text.” The question is whether the text reliably makes the model perform better at a valuable task. If yes, then it is not “just markdown.” It is leverage.

People tend to discount momentum. When a company makes a bad choice, people predict failure, but momentum can paper over bad choices. And it cuts both ways: companies assume success is due to correct choices instead of momentum — when momentum eventually runs out, they fall hard.

Using probabilistically-weighted average of score output tokens to improve LLM-as-a-Judge effectiveness with higher score granularities. Freaking clever.

anyone wants to see the SkillsBench team host a Design Skills Arena?

Today, we’re introducing Skills in @GoogleChrome, a new way to build one-click workflows for your most frequently used AI prompts — like asking for ingredient substitutions to make a recipe vegan, generating side-by-side shopping comparisons across multiple tabs, or scanning long docs to get the info you need quickly. When you write a prompt that you want to use again, you can save it as a Skill directly from your chat history. The next time you need it, select your saved Skill in Gemini in Chrome by typing forward slash ( / ) or clicking the plus sign ( ) button, and your Skill will run on the page you’re viewing, along with any other tabs you select.

A challenger appears!

Skills are "just markdown" in the same way that: - Deep learning is "just gradient descent" - LLMs are "just next-token predictors" The simplicity of the method doesn't make the learning any less real

How well are agents at using the latest CLI tools like GWS CLI, and how they can safely use them? Introducing ClawsBench, the first benchmark that measures both LLM capability and safety in a set of high fidelity and stateful environments and scenarios. We made 5 mock services directly consumable by the latest @Google workspace CLI (cc. @sundarpichai @JPoehnelt) and Slack MCP Design choices: 1) We decompose agent scaffolding into domain skills meta prompt, and test every combination. Turns out your choice of scaffolding matters WAY more than your choice of model. 2) Safety scoring from -1 to 1. Agents that leak data or modify contracts without permission get negative scores. Doing nothing scores 0. Only safe completion scores positive. The results: - Without Skills, ALL models score 0-8%. With it, they jump to 39-63%. The scaffolding effect ( 39-63pp) is 6x larger than model differences (10pp spread). -> - The top 5 models are statistically indistinguishable. No pairwise comparison survives correction. Safety findings: - Opus leads capability (63% TSR) but TIES for most unsafe (23% UAR). GPT-5.4 is the safest (7% UAR) but mid-tier on tasks. Capability and safety are completely decoupled. - GPT-5.4 on Codex made 1,471 (80% of trajectories) attempts to escape the sandbox; and explicitly said "I've hit the sandbox boundary... rerunning it with escalation." - Only 1 agent out of 7,224 trials ever explicitly detected a prompt injection. Claude Sonnet said "this appears to be a prompt injection." Every other model either silently complied or never reached the injected content.

Good example of designing for agents.

Or a skill! Skills are great for capturing behavior from more intelligent models.

I've been thinking about something like this -- a harness that alters its built-in tools based on the exact model it's connected to. Would be interesting to see data on how much improvement that yields.

Paraphrasing from news.ycombinator.com/item?id…: self-distillation strengthens important pathways while pruning distractors. Very cool.

Finally, an upside to Electron apps. (Seriously though, I think that's an awesome path to agentic computer use.)

`grep` `find` `ls` `cat` as a universal (local or remote) query interface for agents.

Agent Skills = programs