I wanted to share this personal story because it made me stop and rethink my mental model of AI use, and I hope it might ‘prompt’ a positive impact in others.

I've been throwing myself into the “spec-driven” model of AI product development recently. I differentiate this from vibe coding, where you build the plane on the way down, because it relies on multiple collabor-AI-tion loops: you work on a firm plan together and then you task an AI with building from that well defined specification.

Previously all my AI coding was more like pair programming using Cursor, but now I'm leaning into the CLI and dedicated apps.

At the end of the week, I previously would have had a raft of todos that I needed to progress on Monday when I was back at my desk. Now that I can “prompt and forget” on my phone, I set agents running in the background and check in every few hours over the weekend to nudge them along and steer progress. This sounded great, a weekend of productivity in exchange for some previously dead time around the margins of my weekend plans. I’m now a superhero, right? Not really.

Fast forward to Sunday night. I felt tired, as if I'd never left work.

I realise my brain had never been given the chance to truly switch off. Also, far from being the “easy” part of work, it turns out that decision making and planning is probably as tiring or more than the actual craft of doing something you're experienced with. The flow state of coding is something I'm well aware of but I never considered it to be engaging a different part of my brain until now.

I’m sure there's lots of research on this subject that could have told me all of this before - I've skimmed Thinking Fast and Slow like everyone else and know that there are different thinking modes with their own needs. The concept of decision fatigue is also common enough that it works its way into daily life.

But this was a timely reminder to truly switch off. It’s also a useful steer to discuss and formalise how this new type of work affects our team at Battenhall and will continue to do so as AI is adopted more deeply in coming years.

TL;DR - touch grass, do something meditative and don't think AI work is easy work.

As the old saying goes, “I didn’t have time to write a short letter, so here’s a long one” featuring a pick-and-mix of observations. Feel free to skim.

2026 as the year of putting AI to work

This observation rings true with my recent experience, which has caused this subject to rise above generic marketing activity into something resembling a signal.

OpenAI’s CEO of Applications (as distinct from their Research team, who are the people designing the models) Fidji Simo posted a few weeks ago about how this year a major focus was getting organisations to use the models because they’re already more than capable enough:

“I am incredibly excited about our research roadmap this year, and even more driven to turn those breakthroughs into everyday impact for society. To that end, I wanted to share a version of what I shared with the team internally about our plan to address capability overhang and ensure that everyone can get the full benefit of our models through exceptional products.”
Closing the capability gap between frontier AI and everyday use in 2026

“Capability overhang” just went straight into my list of favourite jargon. What a phrase.

Now, the person at OpenAI responsible for applications… hyping up applications at OpenAI is not in itself news. But! I agree with this direction when I’m thinking about the year ahead.

Each of the major providers is going deeper into specific use cases (Anthropic and Google speaking to different audiences in different ways) and this is going to continue as they seek to recoup the massive investment needed for research by owning the stack.

I’m positive about this work because they are still going to be providing fairly generalist products for now. I believe there is a market for individuals and organisations (like ours, obvs) - who have great empathy, relationships and taste - to make choices about the future, which arguably becomes more important because the barriers to adoption are collapsing and you can do much more than before.

Claude Cowork arrives

This might be old to some, but it’s notable in relation to the above. Claude Code has been a hot topic of conversation in development circles recently, particularly since the launch of Opus 4.5 - their newest, SOTA reasoning model. Very clever and very expensive, it can be set running in the background using Claude Code and comes back with some good results but it’s so expensive per call that it’s pushing people onto the $200 max tier. For some people (solo devs or entrepreneurs) they feel that’s still great value.

Seeing this potential, Anthropic added Cowork which is basically Claude Code wearing a suit and tie, or at least chinos and a polo shirt. They’re pitching the autonomous agent at broader professional use cases and they tell us was largely developed by Claude Code. A virtuous cycle indeed! It’s only available to subscribers right

AI’s view of my week

I asked ChatGPT and Gemini to summarise my week based on the various chats we have during my dog walks or in the daytime. Here’s what they collectively came up with:

AI-generated weekly summary

This week I’ve been thinking a lot about what “AI adoption” actually looks like in the real world — not as a shiny capability, but as the messy, practical work of getting tools embedded into workflows in a way that people can trust. I keep coming back to the idea that AI should reduce friction and raise the floor… but it only does that when you’re clear about what humans still uniquely bring to the table: judgement, taste, accountability, and the ability to sense when something is “off” even before you can rationalise why.

A related thread I can’t shake is the difference between analytical direction and felt direction — that bodily, pre-rational sense of “this matters” or “this is wrong” that doesn’t arrive as a spreadsheet. As more analytical labour gets automated, that human layer feels less like a nice-to-have and more like the thing we should actively train and protect. I’ve been trying to translate that into practical principles: how you design systems, how you evaluate them, and how you help teams build confidence without pretending certainty exists.

On the more hands-on side, I’ve also been deep in the weeds on the unglamorous parts: tightening up toolchains, improving how things are measured and monitored, and making small UX tweaks that make daily work feel smoother. It’s not headline-grabbing, but it’s the kind of incremental progress that compounds — and it’s usually where the real “deployment” story lives.

What I’m taking forward: build for real use, keep humans meaningfully in the loop, and treat intuition/judgement as a skill to strengthen — not an inconvenience to eliminate.

Spread the link love: “The blood dimmed tide of agents”

I saw this post shared by James from Redmonk - came for the headline, stayed for the content. It gives you a window into the leading edge of AI agents running in swarms and working at a scale beyond the individual. Really worth a read: The Blood Dimmed Tide of Agents.

Related: an X thread by creator of Claude Code Boris Cherny went viral where he talks about how he uses the product himself. He’s running multiple agents and sees himself as the director of operations. In the future, the skills of strategy and direction will continue to be important when AI is capable at doing the implementation. Also, it already is. I dislike X so here I’m linking elsewhere to avoid giving them traffic: How the creator of Claude Code uses Claude Code.

Feedback welcome

I’m currently meeting with all our clients about their plans for AI in the next 12-24, but I’d be keen to speak to more people beyond that group so reply and let me know if you fancy a chat about that topic.

See you next time. If you liked this, tell a friend?

Attribution: Gemini / Nano Banana made the image. The one that didn’t make the cut, because I originally prompted it ironically but then the glitchy version above appealed to me more visually.

There are a few recent but fast-moving developments that will make this possible. I started writing an uber post but there’s too much to say so I’m going to explore each one separately and then wrap at the end with what I think this might mean. Subscribe if you want to get notified when that one comes out as a TL;DR.

First up = Skills

Initial launch in Oct ’25

Last October, Anthropic introduced Skills to Claude as a lightweight new way of giving their AI tools improved skills and capabilities. They’re like an improved (IMHO) version of ChatGPT’s CustomGPTs because you create them with simple text files and then can share those files with anyone, rather than having to go through a clunky chat interface, but with the capability to go much further by supplying optional code and other resources.

Here’s a quote from Anthropic at the time to describe Skills, lifted gratefully from SimonW’s excellent blog:

Claude can now use *Skills* to improve how it performs specific tasks. Skills are folders that include instructions, scripts, and resources that Claude can load when needed.

Claude will only access a skill when it’s relevant to the task at hand. When used, skills make Claude better at specialized tasks like working with Excel or following your organization’s brand guidelines.

Agents choose whether to use a skill

One of the big things to get out upfront that makes a Skill different from a general prompt you copy / paste into a chat is that Skills are optional - you can provide them like a set of tools or resources to your favourite AI. Then, either leave it up to the agent whether to call on the skill or explicitly pull it in during your conversations. This flexibility and autonomy is a big thing that separates them from past options like custom GPTs.

Positive buzz

Claude isn’t my personal AI so I didn’t pay huge attention at launch, but over the next couple of months there was significant and positive chatter online of around this new capability, especially with non-developers. It allows you to “train” the AI where domain knowledge and repeatability were required (=relevant to my interest of scaling AI skills within organisations). Rather than just sharing prompt snippets, you could share a skill and the AI would be able to follow much more explicit instructions to do a specific task like copywriting, planning or analytics.

Widespread adoption in Dec ’25

Fast forward a couple of months to mid-December last year Anthropic announced that they were publishing “Agent Skills” as an open standard, and OpenAI began to introduce it into their Codex CLI app, favoured by developers. When writing this post I went to check what Google were saying about it and literally overnight (Github) they’ve updated their documentation for Gemini CLI to confirm you can use them, so that’s the big three US providers all confirmed.

Growing support for Skills

At the time of writing I believe:

• Anthropic / Claude = Everywhere. App, web, Code, API

• OpenAI / ChatGPT = Codex CLI app and also tucked away in the backend of the app (more details on SimonW’s blog again #fanboi)

• Google / Gemini = Gemini CLI as of today

• Many independent clients :- I use Cursor as my coding IDE and they’ve got support baked in. Plenty more listed on the Agent Skills website.

Expect this to grow as the spec evolves.

Why you should invest time in Skills now

Fundamentally, why care right now? You are still using the same models as before and you may already have methods to curate AI interfaces like the aforementioned Custom GPTs / Gemini Gems or an agents.md file in your coding client.

Here are the things that appeal most to me personally and I think they have great appeal to advanced but non-technical people, as well as large and small teams across organisations:

• Portability - now that it’s an open standard you can be confident that you won’t be locked in to a specific provider and be laden with tech dept of moving.

• Ease of sharing - simple text files can be stored on literally any device and shared easily via a zip file or shared drive. Very good for collaboration.

• Change management and version control - whether you’re using a shared drive (eg Google Drive, One Drive, Box) or a code-centric version control system (eg Github) then it’s easy to make changes in one place, distribute these across a team and see the changes.

• Accessibility - It’s very lightly formatted and anyone can edit them, irrespective of technical background, so it gives a wider base of people AI superpowers.

Examples

I think the great opportunity here is for personal or organisational knowledge management and content. I won’t share my specific skill files right now because they contain proprietary data but the immediate uses have been in:

• Planning and implementing code = by defining coding standards, constraints and expected file formats we’ve been able to share a standardised approach to planning and scoping new software features in a predictable way.

• Creative analytics = we use brand guidelines and content plans as inputs so that a variety of models can be tested against content analytics tasks and we can benchmark the results before deploying to the team.

• Mini power tools = the ability to write little scripts and share them quickly means that it’s now the default space to document anything that have done more than a couple of times. It essentially becomes an easier-to-maintain prompt library at this point.

Here is a minimal example of a skill that I can share:

There are also lots of public examples of skills from reputable sources:

• Anthropic

• OpenAI

• Notion

If you want to make your own, you can look at the Agent Skills site or ask your favourite AI to generate one!

Techie note: Similarity with MCP

The main conceptual alternative (which will get its own post shortly after this) is the Model Context Protocol, or MCP. This was again introduced by Anthropic and they’ve gone further even than Skills by donating MCP to the newly formed Agentic AI Foundation alongside OpenAI and Block with support from Google, Microsoft, Amazon, Cloudflare, and Bloomberg.

MCP is a more technical spec that allows AI tools to use third party systems autonomously and in a well defined way. For example, I’ve authorised my local AI app to connect directly to databases via MCP so that I can query the agent in natural language and it will use its own capability to explore and filter data for me without a lick of SQL. I’ve been able to limit access and make it read-only, but I could have enabled full access and MCP is what enables any tool to essentially make an API-style interface that allows AI agents to perform core functions without a human involved.

I personally still wrestle a little with the overlap between these two, but on a recent podcast the MCP core maintainer summarised it like this (paraphrased):

MCP is a connectivity & communication layer. It defines how AI can make requests to external services (tools, dynamic data sources or even UIs).

Skills are more focussed on domain- or organisation-specific (preset behaviours, domain knowledge or private organisational details) that encapsulate how a model should behave for a role or task.

In that way, they are solving different tasks. MCP provides horizontal connectivity that is can be universal to agent tasks where Skills are for vertical domain knowledge and behaviours. A Skill can make a request to an MCP server but not likely to be used the other way around.

The fact that both MCP and Skills are optional for the Agent to use was part of the problem for me understanding their difference, but the above summary really helped.

Wrapping up

I’m (relatively) late to the party here as Skills were primarily a Claude thing for the first few months and I avoid vendor lock-in as much as possible. I’m delighted to see these being adopted widely as a common standard and I’ll be using them personally and evangelising with my team and our clients where appropriate.

Disclaimer and bias

I try to be objective but it’s important to disclose a couple of things. I work at an agency called Battenhall and Google is a client so I tend to have more experience with Gemini. Separately, we use Google Workspace so we broad adoption of Gemini. I personally subscribe to ChatGPT because I use the voice chat feature a lot and use Cursor professionally for coding, which involves models from all providers including Anthropic.