Re-Analyzing presidents data in Africa

I have had for long this fun idea of writing a good visualization on presidents data and coups in Africa. In fact it was my first post. But I was not so advanced in dataviz up then. I was thinking about updating the visualizations based on web-scraped and available open datasets. Now that we have LLMs, I wanted to explore the quality I can achieve by writing as less a possible code and using an agent (sonnet 4.6 medium). My goal is not really to work on the narrative of issue, this is already well done by other media outlets¹. I want more to explore in which ways our interactions can lead to an interesting data-viz outcome.

This blogpost is the result of my interactions with Claude code, using available tools at my disposal and a bit of coding skills.

This is a long blog post. You should read it as a conversation between me and the agent/assistant, and the outputs are shown here. I was writing and thinking at the same time, so expect some typos in my prompts. Between the lines, I discuss my approach to the problem, and show how time consuming some decisions can be.

You can jump directly to the final output

---

I started by thinking about how to structure the output of our interactions, so that I can have them embedded here in the blogpost.

interaction(c(1, 2))

Here is the output of the .sh file. Notice that LLMs usually use python by default if available (worth learning now the language if you are a die hard R user).

#!/usr/bin/env bash
# Usage: ./log_entry.sh <role> <message>
# role: "user" | "assistant"
# message: the text to log (quote it)

ROLE="$1"
MESSAGE="$2"
LOG_FILE="$(dirname "$0")/conversation.json"

if [[ -z "$ROLE" || -z "$MESSAGE" ]]; then
  echo "Usage: $0 <user|assistant> <message>" >&2
  exit 1
fi

TIMESTAMP="$(date -u +"%Y-%m-%dT%H:%M:%SZ")"

# Bootstrap file if it doesn't exist
if [[ ! -f "$LOG_FILE" ]]; then
  echo '{"conversation": []}' > "$LOG_FILE"
fi

# Append entry using Python (available on macOS without extra deps)
python3 - "$LOG_FILE" "$ROLE" "$TIMESTAMP" "$MESSAGE" <<'PYEOF'
import sys, json

log_file, role, timestamp, message = sys.argv[1], sys.argv[2], sys.argv[3], sys.argv[4]

with open(log_file, "r") as f:
    data = json.load(f)

data["conversation"].append({
    "role": role,
    "timestamp": timestamp,
    "message": message
})

with open(log_file, "w") as f:
    json.dump(data, f, indent=2, ensure_ascii=False)

print(f"[{timestamp}] {role}: logged.")
PYEOF

Even without a full bash or python knowledge, you can guess immediately what the script does. Populate at the end of each interaction the user and assistant answers in a jsonfile. My first instinct was to think: who will run it? before I noticed the agent has already run the ./log_entry.sh at the end of the conversation. So it is updating the file itself. I though using a hook was a less error prone approach but I was quite satisfied with the output.

My next step was to structure how the json will be added into this blogpost.

interaction(c(3, 4))

This is where it got the instructions wrong (or I was not good in writing the prompt). I wanted something like a simple function. You can see this function is named interaction. But I got a file that populates everything at the end of the blogpost. So I asked to correct, and lost some time and money polishing a css (something I could have done myself).

interaction(c(5, 6, 7, 8, 9, 10))

I got the css for the conversations:

<style>
.conv-user,
.conv-assistant {
  margin: 16px 0;
  padding: 10px 16px;
  line-height: 1.6;
}
.conv-user p,
.conv-assistant p,
.conv-user,
.conv-assistant {
  font-family: "Courier New", Courier, monospace;
  font-size: 14px;
}
.conv-user {
  border-left: 5px solid #000;
  background: #fff;
  color: #000;
}
.conv-user p {
  margin: 4px 0;
  color: #000;
}
.conv-assistant {
  border-left: 2px solid #aaa;
  background: #f5f5f5;
  color: #222;
}
.conv-assistant p {
  margin: 4px 0;
  color: #222;
}
.conv-assistant-label {
  font-family: "Courier New", Courier, monospace;
  font-weight: bold;
  font-size: 12px;
  text-transform: uppercase;
  letter-spacing: 0.08em;
  color: #555;
  margin-bottom: 6px;
}
</style>

And the full interaction function I have been using so far:

# conversation_helpers.R
library(jsonlite)
library(knitr)

.conv_data <- local({
  cache <- NULL
  function() {
    if (is.null(cache)) {
      path <- here::here(".draft", "africa-presidents", "conversation.json")
      cache <<- fromJSON(path)$conversation
    }
    cache
  }
})

.render_turn <- function(role, message) {
  css_class <- if (role == "user") "conv-user" else "conv-assistant"

  message <- gsub("\\\\n", "\n", message)
  message <- gsub("&", "&amp;", message, fixed = TRUE)
  message <- gsub("<", "&lt;", message, fixed = TRUE)
  message <- gsub(">", "&gt;", message, fixed = TRUE)

  lines <- strsplit(message, "\n")[[1]]
  paras <- paste0(
    "<p>",
    ifelse(nchar(trimws(lines)) == 0, "&nbsp;", lines),
    "</p>"
  )
  body <- paste(paras, collapse = "\n")

  if (role == "assistant") {
    paste0(
      '<div class="',
      css_class,
      '">',
      '<div class="conv-assistant-label">Claude</div>',
      body,
      '</div>'
    )
  } else {
    paste0('<div class="', css_class, '">', body, '</div>')
  }
}

#' Render one or more conversation turns inline in an Rmd chunk.
#'
#' @param id Integer index (1-based), integer vector, or "all".
#' @examples
#' ```{r, results="asis"}
#' interaction(1)
#' interaction(c(1, 2))
#' interaction("all")
#' ```
interaction <- function(id = "all") {
  conv <- .conv_data()

  idx <- if (identical(id, "all")) {
    seq_len(nrow(conv))
  } else {
    as.integer(id)
  }

  if (any(idx < 1 | idx > nrow(conv))) {
    stop(
      "interaction(): id out of range. Conversation has ",
      nrow(conv),
      " turns."
    )
  }

  html <- paste(
    mapply(.render_turn, conv$role[idx], conv$message[idx]),
    collapse = "\n"
  )
  knitr::asis_output(html)
}

I now had to wire the conversation_helpers.R by sourcing it at the beginning of the Rmd file, after putting it in a correct directory. Claude got this process wrong and my file was not rendering. But I was quickly able to know it came from nested repos and used here to fix it.

I already knew where to download the required dataset for the analyses. So my next prompt was about structuring the required dataset for vizs. I also wanted to use the latest gglite package of Yihuie.

interaction(c(11, 12))

This was a long one. It took quite some times to run, but quite happy about the output:

library(readr)
library(dplyr)
library(lubridate)
library(countrycode)
library(peacesciencer)
library(geodata)
library(sf)
library(here)

OUT <- here("blog-data", "africa-presidents")
dir.create(OUT, showWarnings = FALSE, recursive = TRUE)

# ── 1. Power & Thyne coup dataset ─────────────────────────────────────────────
# Source: https://jonathanmpowell.com/coups/
# coup == 1 → successful coup; coup == 2 → failed attempt
PT_URL <- "https://jonathanmpowell.com/wp-content/uploads/2026/08/pt_20260829.csv"

message("Downloading Powell & Thyne coup data …")
coups_raw <- tryCatch(
  read_csv(PT_URL, show_col_types = FALSE),
  error = function(e) {
    stop(
      "Could not download Powell & Thyne data from:\n  ", PT_URL,
      "\nCheck https://jonathanmpowell.com/coups/ for the current URL."
    )
  }
)

# Standardise column names to lowercase
names(coups_raw) <- tolower(names(coups_raw))

# Add continent + ISO3 via COW numeric code
# The dataset already has a `country` text column; add continent/iso3c via ccode
coups_raw <- coups_raw |>
  mutate(
    iso3c     = countrycode(.data[["ccode"]], "cown", "iso3c"),
    continent = countrycode(.data[["ccode"]], "cown", "continent")
  )

coups_africa <- coups_raw |>
  filter(continent == "Africa") |>
  mutate(
    day         = if ("day" %in% names(coups_raw)) day else 1L,
    date        = as.Date(paste(year, month, day, sep = "-")),
    coup_result = case_when(
      coup == 1 ~ "Success",
      coup == 2 ~ "Attempt",
      TRUE      ~ NA_character_
    )
  ) |>
  select(ccode, iso3c, country, year, month, day, date, coup_result)

message("  African coup events: ", nrow(coups_africa))

# ── 2. Archigos leader tenures ────────────────────────────────────────────────
message("Loading Archigos data from peacesciencer …")
data(archigos)

archigos_clean <- archigos |>
  mutate(
    iso3c     = countrycode(.data[["gwcode"]], "gwn", "iso3c"),
    continent = countrycode(.data[["gwcode"]], "gwn", "continent"),
    country   = countrycode(.data[["gwcode"]], "gwn", "country.name"),
    start     = as.Date(startdate),
    end       = as.Date(enddate)
  )

archigos_africa <- archigos_clean |>
  filter(continent == "Africa") |>
  select(gwcode, iso3c, country, leader, gender, yrborn,
         start, end, entry, exitcode)

message("  African leader tenures: ", nrow(archigos_africa))

# ── 3. Join: who was in power at each coup, who came after ───────────────────
# For each coup event find:
#   - incumbent : leader in power on coup date
#   - successor : next leader after coup (the putschist when entry == "Irregular")

find_leader_on <- function(iso, d, df) {
  rows <- df |>
    filter(iso3c == iso, start <= d, (is.na(end) | end >= d))
  if (nrow(rows) == 0) return(NA_character_)
  rows$leader[1]
}

find_successor <- function(iso, d, df) {
  rows <- df |>
    filter(iso3c == iso, start > d) |>
    arrange(start)
  if (nrow(rows) == 0) return(NA_character_)
  rows$leader[1]
}

message("Linking coup events to Archigos leaders …")
coups_linked <- coups_africa |>
  rowwise() |>
  mutate(
    incumbent = find_leader_on(iso3c, date, archigos_africa),
    successor = find_successor(iso3c, date, archigos_africa)
  ) |>
  ungroup()

# ── 4. Africa map (GADM level 0 = country outlines) ──────────────────────────
message("Downloading GADM Africa map …")
africa_iso3 <- unique(na.omit(coups_africa$iso3c))

# Download each country and bind; geodata caches locally in tempdir by default
gadm_list <- lapply(africa_iso3, function(iso) {
  tryCatch(
    gadm(country = iso, level = 0, path = OUT) |> st_as_sf(),
    error = function(e) {
      message("  Skipping ", iso, ": ", conditionMessage(e))
      NULL
    }
  )
})

africa_map <- bind_rows(Filter(Negate(is.null), gadm_list)) |>
  rename(iso3c = GID_0, country = COUNTRY) |>
  select(iso3c, country, geometry)

# Country centroids for point maps
centroids_sf <- africa_map |> st_make_valid() |> st_centroid()
centroids <- centroids_sf |>
  st_drop_geometry() |>
  mutate(
    lon = st_coordinates(centroids_sf)[, 1],
    lat = st_coordinates(centroids_sf)[, 2]
  ) |>
  select(iso3c, lon, lat)

coups_final <- coups_linked |>
  left_join(centroids, by = "iso3c")

# ── 5. Save ───────────────────────────────────────────────────────────────────
saveRDS(coups_final,    file.path(OUT, "coups_africa.rds"))
saveRDS(archigos_africa, file.path(OUT, "archigos_africa.rds"))
saveRDS(africa_map,     file.path(OUT, "africa_map.rds"))

message("\nSaved to ", OUT, ":")
message("  coups_africa.rds     — ", nrow(coups_final), " coup events")
message("  archigos_africa.rds  — ", nrow(archigos_africa), " leader tenures")
message("  africa_map.rds       — ", nrow(africa_map), " countries (sf)")

Notice that the R script is written also for non interactive session, with messages so a process running Rscript can get informations about what it is doing. This is something I noticed a lot. The LLMs usually write scripts mostly for non-interactive session (although we can use them in interactive). The last part was to layout correctly the specifications of my visualizations, and telling the agent how we want the final output to look like.

Here is my prompt, followed by a lot of refinements I will spare you.

interaction(14)

I think the right approach is to divide my work into small steps again : A prompt for generating the js and the interactions, a prompt for coding the charts with gglite and one prompt for each of the charts. I thought I could send one bulk prompt and get the final output. It failed.

Because I was not having the output I wanted (and was quite frustrated by that), I moved first to fable to set the basis, then came back to opus and sonnet to polish. I added a few more interactions (11 more), after asking fable to fix the output of sonnet, before polishing with opus. I also noticed some datasets were not up to date (archigos stopped in 2015). So I completed with data from wikipedia after confronting the results. I can say that the wikipedia data had a lot of errors, and this crosschecking step surfaced a lot of inconsistencies.

During the interactions, I had crop images from wikipedia of the leaders using opus. I am quite amazed about the quality of the output. Finally I did some few refinements on the labels, data-sources, without touching the code.

Enjoy the final output!²