B Blengi docs
/

Improve your agent

Export a chat for prompt tuning

Any conversation in your dashboard can be downloaded as a prompt-tuning brief — one file containing the exact system prompt your agent runs on plus the full conversation that prompt produced. Paste it into an AI assistant and ask for a revised prompt.

Open any thread under /app/conversations (or an agent's Conversations tab), then use the download icon in the header.

Three ways to take it

  • Copy to clipboard — Markdown, straight into a chat window. The fastest route.
  • Download Markdown (.md) — for anything that renders headings and code blocks.
  • Download plain text (.txt) — identical content with no markup, for editors, tickets or email that mangle Markdown.

Files are named pitchbar-chat-<agent>-<date>-<id>.md so a folder of them stays sortable.

Why the prompt is included

A transcript on its own is not enough to tune a prompt. Pasted alone, an assistant will invent a prompt from scratch instead of revising the real one — it never learns your agent is in Advisor Mode, what your goal note says, or which vertical fragment gets appended at runtime.

So the brief leads with the system prompt as the model actually receives it, rendered through the same builder the live chat uses, with only the retrieved knowledge omitted (those change on every question).

What is in the file

  1. An instruction header telling the assistant what to do with the rest of the file.
  2. Agent settings — name, language, site type, interaction mode, primary goals, strict-knowledge mode and confidence threshold.
  3. The advisor goal note, your own free-text steering, quoted verbatim.
  4. The system prompt actually sent to the model.
  5. The conversation — every turn labelled VISITOR, ASSISTANT, HUMAN OPERATOR or INTERNAL NOTE, with timestamps, the retrieval confidence behind each answer, and the source URLs it cited.
  6. Visitor ratings — any 👍 / 👎 left on a reply (see Response feedback) appears inline with that reply, along with the reason and comment. A reply the visitor marked bad is impossible to miss.

Who can export

Anyone who can open the conversation. This is deliberately not restricted to workspace owners the way the bulk conversation export is — the person tuning a prompt is rarely the person who pays the invoice.

Privacy

The brief contains what visitors typed, which can include names, email addresses or anything else they wrote into the chat. It is the same data the thread page already shows you — but pasting it into a third-party AI assistant sends it outside your company. Read it before you share it, and remove anything personal you don't need.

Nothing is stored: the file is built when you click and streamed straight to your browser. There is no export row, no file on the server and nothing to expire.

A worked example

A common complaint is that the assistant recommends a specific product too early, before it has asked enough. Download the brief for a conversation where that happened, paste it into your AI assistant, and add one line: 

The assistant commits to a specific product before it knows enough
about the visitor. Revise the prompt so it asks about the
differentiating dimension first.

Take the revised wording into Customize appearance → system prompt, or — for Advisor Mode steering — into the agent's advisor goal note, which is usually the faster lever because it does not touch the base prompt.