Asking an AI chat for a table's plan code • rtfreporter

1. What it does

A chat AI writes the R program of one table, from the table’s shell and its ARD: normalize_ard() |> table_plan() |> plan_*(), the document and the file. rtfreporter gives it a kit of facts first, so it writes only functions that exist. Nothing is sent by rtfreporter.

The code here runs on the ADSL that ships with cards (install.packages("cards")). For real work, replace the lines that build the ARD with reading your study’s ARD.

2. Make the kit

library(rtfreporter)
library(cards)

# to try it: an ARD from the ADSL bundled with cards
# real work: ard <- readRDS("the study ARD.rds"), cut to this table's rows
adsl <- ADSL[ADSL$SAFFL == "Y", ]
ard  <- ard_stack(adsl, .by = ARM,
  ard_summary(variables = AGE),
  ard_tabulate(variables = c(AGEGR1, SEX)))
saveRDS(ard, "ard.rds")                         # the same ARD for step 5

rtfreporter_ai_manual("kit/T-14-1-1", ard = ard, overwrite = TRUE)

File	What it is
`00_START_HERE.md`	the request: the program’s three steps, the rules, the answer’s form (one R block)
`01_rtfreporter-ai-user-manual.md`	the manual: the functions that exist (§18), a table from an ARD (§17, §17a)
`02_` … `04_example_*.R`	programs in that layout: ORR by subgroup, AEs by SOC/PT, demographics
`05_ard.csv`	the table’s ARD, normalized (with `ard =`)
`06_ard_shape.md`	the ARD’s summary: the names of its keys, variables and statistics — no numbers (with `ard =`)

The files are written in UTF-8 whatever the locale, on Windows too; Japanese labels survive.

3. What to paste

Paste	Never
the kit’s files `00` to `04` and `06_ard_shape.md` (the ARD’s summary)	subject-level data: ADaM or SDTM datasets, listings, anything with a `USUBJID`
`05_ard.csv` (the ARD’s numbers), only when your rules allow it	other tables’ shells or ARDs
the shell of this table only, as text (rows, columns, cell format, titles, footnotes)

One table a chat.

When the chat takes no attachment, paste PASTE.md: the kit makes it, its files as one text (about 80,000 characters; the ARD’s numbers only with values = TRUE), then the table’s shell under it. Shorter, with the shell in it – the manual’s §0, §17, §17a and §18 and the one example closest to the ARD, about 40,000 characters:

shell <- c("Table 14.1.1  Demographic and Baseline Characteristics",
           "                     Placebo   Xanomeline Low   Xanomeline High",
           "Age (years)  n / Mean (SD) / Median / Min, Max",
           "Sex, n (%)   F / M")
rtfreporter_ai_manual("kit/T-14-1-1", ard = ard, shell = shell, size = "short",
                      overwrite = TRUE)          # or shell = "shell.txt", a file

A shell in Excel: select the whole sheet, copy, paste it into Notepad and save it as UTF-8 (shell.txt), then shell = "shell.txt".

4. How to ask

Write the R program for table T-14-1-1, following 00_START_HERE.md. Answer with one R code block only. The shell is below. The ARD is ard (its shape in 06_ard_shape.md); output_path and report_name exist.

(the shell, pasted)

The answer is one block starting ```r. # QUESTION: lines at its top: answer them in the chat and ask for the program again.

5. Try the code it returns

Save the block’s content as T-14-1-1.R (UTF-8), then run:

library(rtfreporter)
ard         <- readRDS("ard.rds")           # the ARD the kit was made from
output_path <- "output"
report_name <- "T-14-1-1"
dir.create(output_path, showWarnings = FALSE)
source("T-14-1-1.R", encoding = "UTF-8")    # the AI's program
plan_apply(tbl_plan, "table")               # the table as a data frame

It writes output/T-14-1-1.rtf.

What you see	What to do
an error naming a function or an argument	paste the error into the same chat: “this error; correct the program”
`# QUESTION:` lines at the top	answer them in the chat
a row or column missing	check `06_ard_shape.md`: the ARD may not hold it — add it to the ARD, not to the program
the numbers	they are the ARD’s; the plan never computes