AI assistant chat

The AI assistant chat opens a dedicated screen titled Adjust prompt with AI, where you describe what you want to change in natural language, the model proposes a new version of the prompt, and you can iterate as many times as you like before applying.

How it works

1. On the prompt detail page, click AI assistant on either the System prompt or User prompt card.
2. The Adjust prompt with AI screen opens. At the top:
   • Reference product — optional product picker; the chat uses it to render the prompt with real values.
   • Chat model — the AI connection used for the conversation.
3. The body is a three-pane layout:
   • Left pane — chat history with user and assistant bubbles, warnings, and loading state.
   • Bottom of left pane — chat input area with quick preset buttons (Shorter, Longer, Formal, Marketing, etc.) and a multiline text input. Press Enter to send, Shift+Enter for a newline.
   • Right pane — tabbed editor with System prompt, User prompt and Test on product tabs.
4. Each assistant message proposes:
   • An updated system prompt
   • An updated user prompt
   • A short summary of what changed
   • Optional warnings if the model thinks the change might break something
5. The right pane updates automatically with the latest proposal — you can edit either prompt directly.
6. Send another message to refine, click Evaluate to ask the model to critique its own current proposal, or click Apply to copy the draft back into the prompt detail page (you still need to save the prompt to persist).

Reference product

To make the conversation grounded in real data, attach a Reference product. The chat editor uses it to render the prompt with that product’s actual values, so the model sees what the prompt will produce in practice — not the templated placeholders. See Reference product.

Test on product tab

The third tab — Test on product — lets you do a one-shot dry run of the current proposal against a product without applying anything to the underlying template. Pick a reference product (and optionally a target language) and click Run test. The tab shows the rendered output and the token / cost usage of the test call. The fastest way to check whether a refinement holds up before committing it.

Self-evaluation

Click Evaluate to ask the model to critique its current proposal. The warnings array in the response surfaces things the model itself flagged — overly long instructions, contradictory constraints, missing placeholders. Read them.

Suggested adjustments

When the model thinks a runtime parameter (rather than the prompt text) would solve a problem — for example Description: Max length 1500 → 4000 or Name: Min length 800 → 2500 — it surfaces a Suggested adjustments panel under its reply.

• Apply on a single row takes that one suggestion.
• Apply all in the panel header takes every suggestion in the batch.
• Applied rows show Applied so you can tell what’s already in.

These adjustments change the parameters used by the Test on product tab; they do not modify the prompt text. The intent is “before we commit to rewriting the prompt, let’s check whether a different limit / preset would have produced what you wanted”. Run the test again after applying to see the effect.

Sharing test results back to the chat

After Run test, each result card in Previous runs has a Share with AI toggle.

• On (default) — the result is attached to the context of your next chat message, so the model reacts to what the prompt actually produced.
• Off — the result stays in the panel for your reference but is not sent back to the model. Use this for archival runs that would otherwise pollute the conversation.

Each result also carries a Current or Archival badge: Current means the run was produced by the same prompt + parameters you have right now; Archival means the configuration has drifted since.

Comparing the draft to the source

Under each proposed prompt (System prompt / User prompt panes) there’s a Compare with source button. It opens a modal showing the original prompt on the left and the proposed prompt on the right, line by line. Use it on long rewrites where the chat summary alone isn’t enough to tell what changed.

Language

The chat auto-detects the language you write in and replies in the same one. Internally the proposed prompt stays in the original prompt’s language (typically English).

Saved chat sessions

Chats are saved automatically, so you can come back to a refinement later instead of starting over.

• A Chats sidebar lists every saved session, filtered by your admin user and the kind of chat (prompt-template refinement, or the fork wizard’s Discuss with AI entry path). Use the search box to find a session by label.
• Open any session to restore the full message history, the working drafts in the right pane, and the reference product context.
• Rename lets you give a session a meaningful label (the plugin auto-labels new sessions from the first user message).
• Delete removes the session and all its messages.

Saved sessions are personal: each admin only sees their own. The session captures the chat itself — saving the resulting prompt back to the template still requires the Apply button + Save on the prompt detail page.

Tips & gotchas

Tip

Multi-turn refinement is cheap. Iterate three or four times rather than trying to write the perfect instruction up front.

Caution

The chat edits a draft in memory. Until you click Apply and save the prompt detail page, no changes are persisted to the template. The chat itself is saved as you go, so you can re-open it later — but you’ll have to Apply + Save again to push the result into the template.

Related

• Editing prompts
• Reference product
• Test on product