Learning suggestions

Learning suggestions are AI-generated proposals to improve a prompt template, derived from a set of sample products. Instead of asking you to write the perfect prompt up front, the plugin can read a few good examples and suggest how the prompt should be worded to consistently produce that kind of output.

Supported types

In the current version, learning is available for extract_measures only. Other types may be added in future releases.

How to run a learning session

1. Open a prompt template that supports learning.
2. Switch to the Learning Suggestions tab.
3. Click Start learning from samples — the Learn from samples wizard opens.
4. Step 1 — Scope: use the rule builder to pick the products that should be used as samples (filter by active state, categories, manufacturers, etc.).
5. Step 2 — Parameters:
   • Sample size — how many products to process (min 3, max 20).
   • AI connection — which model runs the analysis.
   • Model advisory and estimated-cost hints appear inline.
6. Step 3 — Review: summary of template name, scope label, sample size, plus a preview of the first five matching products.
7. Click Start & watch (or Run in background) — the wizard dispatches a learn_prompt job and closes.

The job runs asynchronously. The Learning Suggestions tab polls every few seconds and shows progress while it runs.

Reviewing suggestions

When the job finishes, each suggestion appears as a row with:

• Status — pending / applied / discarded / failed
• Confidence — per-suggestion score
• Created — timestamp

Per-row actions:

• Review — opens the Review learning suggestion screen with tabs for System prompt, User prompt, Samples, Rationale. You can edit before applying.
• Apply directly — merges the suggestion into the template and marks the row as applied.
• Apply as new fork… — opens the Fork wizard pre-filled with the suggestion.
• Discard — keeps the row for audit but does not apply.
• Delete — removes the row entirely.

Tips & gotchas

Tip

Pick samples that exercise different shapes of input. If all your samples have the same description structure, the suggested prompt will only be tuned for that shape.

Caution

Apply directly modifies the template in place. If the change is undesirable, the prompt detail page’s reset-to-default action still works, but you’ll lose any other edits you made on top. For risky changes, use Apply as new fork… instead.

Related

• Editing prompts
• Forks — scope a learned variant rather than replacing the default