Using alongside an LLM (AI)

NOTE

When working with Large Language Models (LLMs) consider your organization’s guidelines as well as any laws regarding the usage of epidemic data.

Online LLMs (ChatGPT, Claude, Gemini) usually use your data to train their models so a version of your data might become completely public. Prompt them to help you create the nowcasting models but do not pass them real data unless you are absolutely certain of what you are doing (both technically and legally).

Local LLMs (Llama, Mistral, Qwen) can be run inside your computer without the data ever leaving your system. They are less powerful but help with handling data privately

Pair-programming with an LLM (`SKILL.md`)

diseasenowcasting ships a reference written for LLM assistants: SKILL.md.

This recipe is the same for any LLM. Put the contents of SKILL.md into the assistant’s context, then prompt:

Get the file. Open the link above and copy the text. Or download using R:

skill_url <- "https://raw.githubusercontent.com/RodrigoZepeda/diseasenowcasting/master/SKILL.md"
download.file(
  url = skill_url,
  destfile = "SKILL.md",
  method = "libcurl" 
)

Hand it to the LLM
- Claude (claude.ai or Claude Code): attach SKILL.md or paste it. In Claude Code, drop it into the project and it loads as a skill automatically.
- ChatGPT: paste SKILL.md into a Project / custom-GPT instruction, or attach the file to the conversation.
- Gemini: paste SKILL.md into the prompt (or a Gem’s instructions).
- Local models (Llama, Mistral, Qwen …): put SKILL.md in the system prompt. If the context window is small, paste only the sections you need (e.g. “2. The model menu”, “3. Data preparation”, “4. Fitting”).

Prompt. For example using ChatGPT one can pass the prompt along the file:

The file SKILL.md is an API reference for the `R` package
 `diseasenowcasting`. Using this API write R code that builds a 
 tbl_now from a weekly linelist with columns `onset`, `reported`, 
 and `region`, gits an HSGP with a negative binomial likelihood and a 
 generalized gamma delay and plot the results.

And get as a response:

<THIS ANSWER COMES FROM CHATGPT'S MODEL GPT-5.5 INSTANT AS OF JUNE 8TH 2026> 
Based solely on the API in SKILL.md, the workflow is:

1. Build a tbl_now from the weekly linelist.
2. Define a model with:
    `nb_likelihood()`
    `hsgp_epidemic()`
    `generalized_gamma_delay()`
3. Fit with `nowcast()`.
4. Generate predictions with `predict()`.
5. Plot with `autoplot()`.

```R
library(diseasenowcasting)
library(tbl.now)
library(ggplot2)

# weekly linelist with columns:
# onset    : event date
# reported : report date
# region   : geographic stratum

# Build tbl_now
tn <- tbl_now(
  weekly_linelist,
  event_date = onset,
  report_date = reported,
  strata = region,
  data_type = "linelist",
  verbose = FALSE
)

# HSGP epidemic + Negative Binomial likelihood +
# Generalized Gamma reporting delay
mdl <- model(
  nb_likelihood(),
  hsgp_epidemic(),
  generalized_gamma_delay()
)

# Fit nowcast model
nc <- nowcast(
  data = tn,
  model = mdl,
  n_draws = 2000
)

# Posterior predictive nowcast
pred <- predict(nc)

# Plot median and uncertainty ribbons
autoplot(pred)

```

diseasenowcasting team

Pair-programming with an LLM (SKILL.md)

Pair-programming with an LLM (`SKILL.md`)