Trigger Mode

Description tuning now lives under Eval Workbench → Trigger.

Use Trigger mode when a skill fires too often, misses requests it should catch, or needs a tighter trigger boundary.

Use it when a skill is triggered too often, not triggered when it should be, or has a description that is too broad or too vague.

What's on this screen

The mode has three main areas:

• Trigger Eval Queries — example requests split into Should Trigger and Should Not Trigger
• Trigger comparison — runs the candidate comparison loop
• Results — shows score progression, the before/after description, and actions for the best result

Generate trigger eval queries

1. Open a skill and switch to Eval Workbench.
2. Select Trigger.
3. Click Generate candidates.

Generated candidates appear under the baseline description so you can compare wording before running the trigger pass.

Edit trigger cases

Each query belongs in one of two columns:

• Should Trigger — requests that should invoke the skill
• Should Not Trigger — requests that should not invoke the skill

You can edit each saved case directly, switch Should trigger on or off, add cases, and delete cases before saving the prompt set.

Run trigger comparison

1. Make sure at least one query is in Should Trigger.
2. Click Run comparison.
3. Review the baseline plus generated candidates after the run completes.

Click Cancel to stop an active comparison run.

If every case is marked as not triggering, the mode cannot produce a useful trigger comparison.

Apply the recommended description

When comparison completes, the mode shows:

• the baseline description
• generated candidates
• pass summaries from the completed run
• a Recommended marker on the strongest candidate

Click Apply on a candidate card to save that description.

What you'll see

• Empty query state — No queries yet. Generate or add them manually.
• Generation running — candidate generation is in progress
• Comparison running — a comparison run is in progress with progress text and Cancel
• Navigation guard — Process Running with Stay and Leave when you try to leave during an active run

Quick reference

Control	What it does
Generate candidates	Creates trigger-description alternatives from the current prompt set
Run comparison	Compares the baseline plus generated candidates
Cancel	Stops an active comparison run
Apply	Saves a selected description candidate