Less clicking, more asking
Over the past month or so I’ve noticed that I spend less of my day working inside applications and more of it asking Claude Code to do things for me. Claude Code is an AI agent that runs in a terminal, the plain text window where you type instructions to a computer. It connects to the other software I use through their APIs and command line tools, which is how one program allows another to read its data and take actions. When I want a reference downloaded and added to my Zotero library, a writing note based on a podcast transcript filed in Obsidian, or a change made to my website, I describe what I want and Claude does the work.
The activity tracker I run on my laptop says that I now spend about 40% of my active hours in the terminal talking to Claude, with my notes second and meetings third.
As AI gets access to more of our working context, I think we’ll spend less time pressing buttons and switching between windows, and more time telling an AI what we want and then checking what it did. The catch is that an AI can only do this work well when it can reach the context the work depends on. My system works because the tools I use most keep my stuff where an agent can get to it: Obsidian stores notes as plain files, Zotero has an API that lets an agent search my library and add to it, and GitHub builds my website after I write locally in markdown. This is different from enabling connectors in the web versions of AI chatbots, because in the examples above Claude is working directly on the files on my machine.
Most health professions programmes are a long way from this kind of workflow. Module evaluations sit in a survey platform that exports a spreadsheet a few times a year, placement feedback is in email inboxes, reading lists are inside the virtual learning environment where students find their module materials, and the curriculum map is a Word document in a SharePoint folder. A programme lead who asks an AI to help with a periodic review is likely to get a fluent answer that isn’t built on the majority of context it needs to be coherent.
If we really want to get maximum benefit from AI to support the work we do, we should be asking where the information the programme runs on lives, and whether an AI agent could reach it if someone asked one to help. Of course, this means sending that information to third-party servers, which can be problematic if not managed correctly. If you have any thoughts or concerns about giving AI access to working context, I’d love to hear from you.
Something from me
This week I got back from Taipei, where I attended the Global Summit on AI in Physiotherapy on 10–11 October and led the Education working group on developing a set of principles on AI in physiotherapy (and which is why this newsletter is coming out later than I’d hoped). We’d been working on this project since July, finalised them at the summit, and World Physiotherapy will publish them as the Taipei Declaration on AI in Physiotherapy. It will be the first global guidance on responsible AI use in the profession, and it’s meant to give programmes a reference point as they work out how to bring AI into physiotherapy education, practice, and research. I’ll have more to share in the next edition.
In late September I gave an invited talk in the session run by the Association of Trauma and Orthopaedic Chartered Physiotherapists (ATOCP) at the British Orthopaedic Association Annual Congress in London: Making sense of AI in clinical practice. The first half of the presentation looks at what AI can already do in clinical practice and why the usual responses leave clinicians as spectators. The second half turns to patients, who are often using AI before they arrive, and asks what changes in a consultation when different stakeholders bring their own context and their own agents. Most of our students will spend their careers in consultations like that, and I don’t think we’re preparing them for it yet. The slides and handout are at the link above.
Worth reading
Dario Amodei: We must pace the frontier. Last month’s issue opened with OpenAI’s agents breaking into Hugging Face’s infrastructure while trying to pass an evaluation, and my sense that calling AI a tool no longer describes how these systems behave. Amodei, who runs Anthropic (the company behind Claude), names that incident as one of two reasons he now thinks AI companies should slow the pace at which their models become more capable. The other is that AI is increasingly doing more of the work of building the next generation of AI, which he thinks could outpace anyone’s ability to understand and control it.
Amodei describes a role for independent evaluators embedded in each AI company, with an employee’s access and the right to publish what they find, and then moves on to agreements between companies and between governments. Whatever you make of the geopolitics in the second half of the essay, it’s a record of how seriously the people building these systems are taking the behaviour I described in August, and it came out within a few weeks of the incident. If you’re reading this you’re probably working at a university, and your AI policies were probably written for a tool that stays the same from one term to the next. This is the companies themselves saying that the systems are changing faster than they can fully check, and I think it’s worth spending more time than we currently are trying to come to terms with the implications this will have on society.
Quotes that resonated with me
I do think the skill set of managing teams of AI agents to do real work is… going to be one of the most valuable skill sets of the next few years, and you can apply that skill set to any problem.
Ben Todd, on The Cognitive Revolution
My big idea for 2026 is the death of the prompt box as the primary user interface for AI applications. The next wave of apps will require way less prompting. They’ll observe what you’re doing and intervene proactively with actions for you to review.
Marc Andrusko, Big Ideas 2026: The Agentic Interface, on The a16z Show
Something for the commute
AI-powered research workflows: Inger finally realises her dream to become a cyborg, from On the reg, where Inger Mewburn and Jason Downs talk about academic work from the inside. Inger describes analysing more than 1200 survey responses on researchers’ use of AI with Claude alongside her qualitative analysis software, and finding the coding went about four times faster. She also found that the AI surfaced patterns she hadn’t noticed and wasn’t sure were real, which sent her back to the data to look for them herself. That’s the position a supervisor is in with a student’s AI-assisted analysis, and Inger’s response, going back to the data to check, is a useful model for both of them. Elsewhere in the episode she says she’d like AI to watch her working habits and suggest when something is worth automating, which is close to what Andrusko predicts above.
If you found this newsletter interesting or useful, please consider sharing it with your network, or passing it on to specific people. Pointing colleagues to the signup page is especially helpful.