Prompts

The brief that made this site, and the steers that followed

This site was built with coding agents, working from a brief and a run of corrections. This page is what I asked for and what I said back. The JuliaCon 2026 site did the same, and the how I LLM deck before it put its prompt on a slide.

The brief

This is my opening prompt, rewritten as plain prose and grouped under headings. The spelling and grammar are corrected and the sentences are complete. The judgements are unchanged and nothing has been added. The verbatim version, typos and all, is in the repo. It began with the workshop programme and the organisers’ request, which are on the talk pages.

The site

I am not yet sure what to put in these talks. To begin, use agents to create a folder in seabbs called AIandIDVUB, with the same structure as the one I made for JuliaCon 2026. Put the slide decks in place with their titles and QR codes, but leave the content for now. I will then make a matching GitHub repository and give you access to push to it. I think the abstract for each talk can be the organisers’ request, as bullets.

The research

Use agents to search my recent work for anything related. I am unsure whether the organisers mean agentic AI, deep learning, universal differential equations, or something else.

The keynote

Before I looked back at what the workshop covers, I had planned the first talk in a few sections. The first would be a ten minute introduction to outbreak modelling over the ten years or so I have been in the field. It would cover Ebola in West Africa, which I was aware of and others worked on, then COVID-19, mpox and the current Ebola outbreak. The second would be ten minutes on agentic AI. I have material on this from how I LLM, but I also want to ask what it means for modelling. The third might be ten minutes on the opportunities in other AI models. Universal differential equations, physics-informed neural networks and neural flows for fitting, on which the nfidd course had a little, at a high level. Perhaps foundation models for outbreaks and infectious disease forecasting, spun out of the current time series foundation models. Perhaps agentic model composition, such as Google’s forecasting success at FluSight.

The communication lecture

For the communication lecture I am not sure what to do. Mine my recent work and give me a proposal.

How it should look

Look at the git history of my recent talks to see how I have pruned and refined the language and style to make them my own, and follow that. Be wary of dense text on slides. We like schematics, pictures and sparse, clean slides.

The steers

Follow-up instructions during the build, in order, rewritten in the same way. The two entries in square brackets are notes on what I forwarded rather than things I said. The running version is in the repo.

Present the prompt log as we did for JuliaCon, cleaned up in the same way, rather than as given here.

A reminder that you can use a workflow and subagents. Structure, research and content can run in parallel, with cross review.

You should now have an invitation to the repository for pushing.

You can push to main and iterate.

A reminder to update the remote repository as you go.

Please have agents start drafting the content and cross review it against my specification. The Google point was about LLM-guided tree search. Note the need for pictures, interactives and the like.

[Forwarded the organiser’s practical email for the research exchange on 15 September, having asked whether there was a theme for the day.]

For the exchange I might like 20 minutes on composable EpiAware modelling and its potential links to AI approaches as this group might think of them, such as composition, universal differential equations and physics-informed neural networks. There may be nothing doing there at the moment. Then ten minutes on the recent delays work and epidist, where the AI cross links are less clear. Then perhaps five minutes on the workflow and where AI sits. I am not sure. Review this and come up with a first version after cross review.

[The organiser replied that there is no theme for the exchange, just an introduction of my work and group.]

So I think the second day needs a bit on my epinowcast organisation and its seminar, a loose forum with users, and Clara and Fanny organising the seminar. Also epiforecasts, led by Seb Funk, which I am associated with. The cross-over is not clear-cut, but it is real-time modelling and evaluating it. CMMID …

Where it makes sense, mine my past presentations heavily for images, complete slides and sections.

Use an agent to make the prompts on the site a bit more formal and less confusing. Do the same for the abstracts, more formal but steered by what the organisers said.

You can use less powerful agents where that makes sense for a task.

Can we look at my recent packages and papers and get a slide with collaborators, or a few slides organised in some logical way? Use a lower powered series of subagents.

How it was built

Five research agents ran first, before any slide was written, and wrote their notes into notes/.

• what I have already said about coding agents, and the live DRC outbreak model built with them
• what my teaching and grants say about the other AI methods, and what the literature says about foundation models, universal differential equations, and flows in epidemiology
• my record of communicating with government, the public, and other scientists during COVID-19
• what the other speakers work on, and so what this workshop means by AI
• a timeline of the outbreaks I have worked on or watched

The style notes were carried over from the JuliaCon 2026 build, including the catalogue of LLM phrasing I strip out of my own slides. The prompts on this page and the abstracts on the talk pages were rewritten by an agent, on the steer above. The commit history is the record of what happened next.