Coding agents built the site and drafted the slides, working from a brief and a long run of corrections. This page is what I asked for and what I said back. The how I LLM deck put its prompt on a slide at the end. There are three decks this time, so the prompts live here and each deck links back to this page instead.
The brief
This is my opening prompt, cleaned up for spelling and sentence boundaries and grouped under headings. The judgements are unchanged. The verbatim version, typos and all, is in the repo.
The site
I have made a new repo, JuliaCon2026, and invited seabbs-bot to it. I need to give three presentations with three posters this week at JuliaCon. What I want is a simple home page in Quarto, in the style of my personal seabbs.github.io site, with a page per project that has the title and the abstract, a link to the slides, and a bullet list of resources that are linked. I then want a home page that has a bit about me with a link to my website, and some links to my other recent interesting work that a JuliaCon user might like.
I want the QR code on the slides to go to the home page of the project on our little website, not to the slides this time.
The delays deck
I want to make the slide decks for each talk in Quarto again, matching how I did it for my recent how I LLM talk (see that repo for its prompt and my follow-up prompts) and my how to serial interval talk.
For the CensoredDistributions talk I want to show some of the primarycensored code, using Quarto’s support for that, and again the Julia code to compare. It is not a straight win. primarycensored has over 10,000 downloads and is used in an ecosystem of packages by public health departments. CensoredDistributions is used currently by me and a few close collaborators. We had to specify the problem as an ODE in Stan, but otherwise it just worked with AD. In Julia it was quite painful (see the repo git history).
Note this is going in the Julia direction of being modularised, into ConvolvedDistributions and more work on expanding distributions and estimating complex compound distributions like the generation time.
The roadmap deck
For the roadmap and org slides I want the why. epinowcast (the “why am I so late” slides might have good context for this, as does the compose grant), EpiNow2’s lack of expandability and modularity, reference to other projects. Multiple ways to do it, via ComposableTuringIDModels, via distributions, others.
Julia standards by default are somewhat sloppy and you need to enforce a template. Most of the templates are fairly modular; I want it to be rigid, to enforce uniformity. We needed to set up AD testing, as you cannot say which AD backend is best. They are all flaky and dodgy, or run by small teams. No robust tooling for this. DifferentiationInterfaceTest is very nice but needed more on top. The lack of standards for this in wider Julia means making a model is basically always a custom affair with lots of iteration. That is not realistic for most infectious disease modelling, and especially not in an outbreak or another new setting.
Very heavy LLM use for the org, with max Claude driving it. Pull in the PR figure from how I LLM. What does this mean for contributions? Flag the Chris R keynote or long talk on this. Do we build out an org agent which acts in batch overnight against public standards? Where does it run safely?
For the org talk I have a bunch of questions. What does a good Julia org look like? How do you depend on packages in the ecosystem when maintenance is usually fairly minimal for many packages? What does treating AD as a first class citizen look like? Are there examples of Julia ecosystems with broad adoption? SciML maybe, but my impression is very technical folk. SpeedyWeather maybe? Some of the physics ones, not sure. Try and use my recent work and Claude history to understand what other questions I might have.
The composability deck
For the case for composability we have a bunch of talks in that private repo we can repurpose and adapt, but they need to get down to 15 minutes, and not the Turing 15 minute edit, as that was overly Turing focused. Since writing that paper I have become a lot less sure about Turing as a sensible backend, due to reduction in team size, lack of stability, and not complete AD backend support. I am now exploring composable distributions and similar. Bring in some of the other talks.
How they should look, and how they should be made
Have a page on the JuliaCon 2026 site we are building that is the prompts. Clean this up and give it some structure and have it there, and then a steering section, and then we link to it from the slides, rather than embedding it in the slides as we do in how I LLM. Make sure to get my prompt and steering prompt from there.
We want a large workflow cross checking and exploring my repos, agents on core building, pushing to main is okay for this repo. Agents cross checking slide length and quality as well as correctness. No stuff about my citations.
Lots of work on this is okay, and in fact good, as we want it to be good, clear, concise, with nice use of bullets. Similar to previous talks, and in a similar voice to my blog and past presentations.
We have a poster for each of these but I was thinking that we would print the QR slide, the title slide, and two others. Those two should be fancy/interesting. Perhaps that has its own bullet, summary slides or similar, that is just those.
What I am unsure about
Something I am a little unsure about is how and where to bring in my recent BVDOutbreakSize work. I think that maybe this could be in the composable org part, as to where we are, a bunch of LLM driven blah, very long, hard to check, hard to steer. But on the same front it is much better in my opinion than what we had before. It can consider lots of datasets, adapt as data changes, read sitreps in French, and scan out data from images with little effort ourselves. It is easier to follow our workflow, which is a major motivator for the composable org.
The steers
Follow-up instructions during the build, verbatim apart from spelling. There are thirty-one so far and the build is not finished. It is appended to as they arrive, and the running version is in the repo. Seven are about the workflow, twenty about the content, and four about the layout. A few about specific third party packages are left out.
Workflow
Please make sure you get this website up and building and iterate on it. Keep working and review using Opus agents against my prompt above to make sure it is hitting what I wanted.
I asked for a workflow with heavy agent usage, an Opus and Sonnet mixture.
Where are the searching context pages and the workflow?
We need another review stage after the fix to confirm and potentially loop.
That is kind of what we want I think.
Make sure workflow members are reviewing and improving the site as well as just the slides.
Reminder to update the prompts page, but do not offend anyone.
Content
Other recent work might want to link to i.e epinowcast?
No, the point was an LLM wrote the outline of me and I wanted that to be clear.
It is not really part of what these talks are about, cut that. There is less LLM emphasis here than we had in how I LLM, less self referential, but we do discuss as in the original prompt.
We already have a robot opener so we do not need the closing one at the end at all I think.
The following was drafted by @seabbs-bot who has a very suspiciously high opinion of me so take it with a pinch of salt.
Link to seabbs-bot GitHub profile.
I think this one is on Friday? [the composable talk, 2026-08-14 16:45-17:00]
We need the QRs in the second slide of each talk with bullets outlining what we talk about. We do not need the QRs on the home page for each project as they are already on it. Not all the text was generated by LLM, remove that. The point we were trying to make was the bio was LLM generated. Get agents to scan the projects and talks for LLM-isms we removed or asked to remove in our other presentations (get them to look at the git commit history) and clean them up. “Something something is the point” or “sentence, some random point” are both common patterns. Make sure slides are visually interesting. The primarycensored talk needs a second slide on right truncation after double interval censoring.
Also the composability paper has weird EpiNow2 justification vibes in the slides we have here that were not in the original. Also we now have ComposableTuringIDModels which is the prototype we talk about in those slides. It is not linked from the project page or otherwise mentioned clearly but should be.
Yes, everything in the 40 minute talk is canonical.
The slides honestly kind of suck. In general the content is not my style. How I LLM is not a good reference as it was all bots; the other talks are a better reference. The new diagrams are high level and not very good, and there is lots of text all over the place. Slides have become very cluttered.
For the delays talk we need what an epidemiological delay is, biases, double censoring, truncation, with a diagram as we had in the recent serial interval talk. Then the software for this. primarycensored, what is it. Then we need partial pooling, so how do we extend it, brms, okay, epidist, but. Then what about the population level, nowcasting, epinowcast. Then what about in Julia, CensoredDistributions.
For the ecosystem talk it is over indexed on the R stuff. It should instead be that we want to build this composable thing, then what we wanted: something like epinowcast, something like SciML, something like the Turing community. But that approach was hard, and multiple approaches are possible so I want to leave space for that. Something like rOpenSci for epi, ideally.
We can still have what we needed to make it work for fitting. Stan is what is used in epidist and epinowcast.
“Stan is the through line” is the kind of classic LLM language that I do not use and edit out.
brms is a nice way to metaprogram Stan but is limited in terms of being extensible. In epidist we try hard to extend it but it took a lot of effort. If Julia had something like brms, which a few people, maybe here, have been talking about, it would very likely be a lot more extensible and flexible.
We might not need one for a lot of use cases. People can more easily reuse CensoredDistributions in any model.
This from the delays project page is a classic LLMism I would not use in my own writing: “The switch to a fixed rule, and what it cost”.
The paper being in CDC clearance is the bullet. ComposableTuringIDModels and epiaware.org have most of what we need, and the latter has an approaches page we can link to with the requirements.
Layout
Put this in italics [the bio], from @seabbs-bot with robot emoji or similar. Bit of fun, breaks it up for being a bit pompous.
The home page is long and kind of empty, can we make it wider or move the projects into a line with less length or similar compression to make it more interesting?
It can be a little busier, let us not be so binary.
On the pages, put related talks under the abstract to make better use of the space.
Seven of the thirty-one are corrections to the workflow rather than to the content. The brief said what the talks should say. It did not say how hard to look before saying it. The content steers were mostly about what to cut.
How it was built
Nine research agents ran first, one per question, before any slide was written. They wrote about 3,500 lines of notes into notes/.
- the git history and issue trail of primarycensored and CensoredDistributions.jl, for what actually went wrong in Stan and in Julia
- side by side code snippets in Stan, R, and Julia, each one run to check it works
- download counts and reverse dependencies for primarycensored and CensoredDistributions.jl
- the EpiAware organisation itself, its package template, its AD testing, and its CI
- why the previous attempts at modularity did not work, from epinowcast and EpiNow2
- agent use across the organisation, and what it has done to pull requests
- the existing composability decks, and what to cut to reach 15 minutes
- how SciML, JuMP, and Turing govern themselves, for the questions at the end
- the BVDOutbreakSize model, and where a live outbreak analysis fits in
They were told never to state a number they had not checked against a repo, the GitHub API, or a package registry, and to list separately anything they could not verify. Several notes have an unverified section. Those claims are not on the slides.
Some of what came back contradicted the brief. The brief says the Stan side needed an ODE but otherwise AD just worked. The Julia history says both languages moved off adaptive quadrature by default because the differentiator could not follow it, and that 48 of the 396 issues in CensoredDistributions.jl have AD, autodiff, gradient, numerical, quadrature, precision or performance in the title. So the deck says that. One agent was asked for a lines of code comparison between the two packages. It counted the lines in both, then said the number should not go on a slide because the scopes are not the same. Another declined to claim public health agency use of primarycensored, because the only evidence was the epinowcast team describing its own tools to a public health audience.
Writing agents then drafted the slide text from the notes. Review agents read the drafts back against this brief for length, for voice, and for claims that were not in the notes. This page was written while that was still going on.