Blog
Notes & Updates
New developments, useful finds, and things worth knowing about AI for public health. Posted when something is worth sharing — not on a schedule.
Blog
New developments, useful finds, and things worth knowing about AI for public health. Posted when something is worth sharing — not on a schedule.
If you work for a national public health agency, a ministry, an international health organisation, or any public institution that runs Microsoft 365, there is a reasonable chance Microsoft Copilot is already available to you through your employer's subscription at no additional cost and with no separate sign-up required.
If that is the case, then Copilot is sitting inside the tools you already open every day: Word, Excel, Outlook, and Teams. You do not need to install anything or create a new account; you just need to know where to look.
The practical difference this makes is significant. Working in an institutional setting often means handling documents that should not leave your organisation's data environment — draft policies, internal briefings, personnel matters, surveillance data that has not yet been published. Consumer AI tools process your input on external servers; enterprise Copilot processes it within your organisation's Microsoft 365 boundary, subject to the same data governance policies as your email and SharePoint. For a lot of public health work, that distinction matters.
The Copilot page on this site explains the difference between the consumer and enterprise versions, what Copilot can do for public health tasks specifically, and how to check whether your organisation has it enabled. If you work in an institution and have never looked into this, it is worth five minutes of your time.
One rule worth holding to, whichever tool you use: never upload work documents into your personal Cowork setup. Take some time to familiarise yourself with your organisation's responsible AI use policy, if it has one, before putting any work content into any AI tool.
For more on the responsible use of AI in public health, these two reports are worth reading: Responsible artificial intelligence in public health: a Delphi study on risk communication, community engagement and infodemic management (BMJ Global Health, 2025) and a World Health Organization report on AI in health.
The Health Campaign Builder is now available in the repository. It guides Claude through a series of questions about your target population, languages, reading level, and visual theme before generating a complete, accessible HTML health page. This is how it came to exist.
I used the skill creator — a tool built into Claude Cowork. I described the problem I was trying to solve: producing tailored health communication materials for specific populations takes time and expertise that most public health teams do not have to spare. I wanted a skill that would guide the thinking upfront and produce a finished, ready-to-use page in one session. Claude drafted the skill from my description. The skill asks eight questions before generating anything: health topic, target population, key message, languages, format, visual theme, branding, and reading level. Only after I confirm all of these does it produce output.
To test whether the skill made a real difference, Claude ran six agents simultaneously — three scenarios, each run twice in parallel, once using the skill and once without it. Running them side by side meant the comparison was honest: both worked from the same prompt under the same conditions, so any difference came from the skill itself. The three scenarios were a cervical screening page for older Maltese-speaking women, a mental health resource for young people, and a multilingual TB screening page for migrants in Malta in English, Arabic, and French. Claude collected the outputs into a review viewer and asked me to evaluate all six pages.
The with-skill outputs showed more deliberate calibration — appropriate colour choices, adjusted reading levels, right-to-left support for Arabic. The without-skill outputs produced functional HTML but missed the population-specific details that matter in practice. I gave feedback: the mental health page used too bright an orange; the Maltese translation was poor and the font did not support Maltese special characters. Claude updated the skill accordingly — a calmer colour palette for mental health themes, Noto Sans as the default font for Maltese, and a visible warning on any page where minority language translation quality may be unreliable. The whole process took one session.
The skill is available in the repository.
When I built the original version of this site, I wrote separate setup guides for Claude (browser), Gemini, and ChatGPT. It made sense at the time — but in practice, the core setup steps are nearly identical across all three platforms. The meaningful differences come down to where to find the settings and what they are called, not the underlying approach.
I have merged those three pages into one: Other platforms. It covers Claude in the browser, Gemini (and Gemini Spark), ChatGPT and Codex, and Perplexity Computer, with the essential steps for each and direct links to the providers' own documentation for anything more detailed. The principle is that a good setup guide does not try to compete with the platform's official docs — it tells you what to do and points you where to go for more.
I also added a dedicated page for Microsoft Copilot. Most public health professionals working in institutional settings will have access to it through their organisation's Microsoft 365 subscription, often without realising it. The Copilot Notebooks feature is particularly useful for research synthesis, and the enterprise data protection makes it the safest option for working with organisational documents. It seemed wrong not to cover it.
The Claude Cowork setup guide remains the main guide on this site — it is the platform I know best and the one I would recommend for getting the most out of agentic AI in public health work. But the right tool depends on your situation, and that is now easier to navigate from the front page.