37 Chat: multi-turn conversations

So far we’ve sent one prompt and got one answer back. Often you want an ongoing conversation — to problem-solve, refine an answer, or just chat. This is easy with ellmer (fiddlier with plain Python or base R, so we’ll stick to ellmer here).

library(ellmer)

strong_model <- "anthropic/claude-sonnet-4"

chat <- chat_openrouter(
  system_prompt = "You are a helpful assistant who specialises in ecological statistics.",
  model = strong_model,
  api_args = list(max_tokens = 2000)
)

live_console(chat)
# live_browser(chat)

live_console() lets you chat right in the terminal. live_browser() opens a local browser window that feels a lot like ChatGPT’s interface.

LLMs perform better when you put everything in one well-prepared initial prompt, rather than drip-feeding it over a multi-turn conversation. Responses to a problem split across multiple questions are less accurate than the same information given up front (Laban et al. 2025) — this holds across logic, coding and everything in between.

This is the same “context poisoning” problem we ran into with agents in Section 2: once a wrong answer is sitting in the chat history, it’s hard to get the model to un-learn it within that thread. If a multi-turn conversation goes off the rails, the fix usually isn’t to keep correcting it — it’s to start a new chat with a clearer, more complete initial prompt.

That said, it’s fine to use a chat as a genuine problem-solving partner when you’re not sure what you want yet. Just don’t mistake “figuring it out together” for “getting the best possible answer” — once you’ve got a plan, write one clear, detailed prompt to get the final version.

Challenge

Start a live_console() chat and ask “How could I test if fish abundance depends on coral cover?” over four separate turns (one clarifying detail per turn: response variable, sampling method, number of sites, that it’s count data). Then start a fresh chat and ask the same thing as one single, fully detailed prompt. Compare the two final answers.