Combining answers

A merge is one more chat request over the answers a Beam run collected. It goes to a model you pick, not necessarily one that answered. Every mode sends the same four things, in this order:

system       the mode's system prompt
user/asst    the whole conversation, again
assistant    each answer, one message per card, in card order
user         the mode's merge instruction

The four modes differ in those two prompts, and in how many steps they run.

The four modes

Mode	Button	What comes out	Requests	Use it for
Fuse (default)	Add Fusion	One merged answer, streamed into the card	1	Work where the union of the findings is the deliverable: a code review, a bug hunt, a draft.
Guided	Add Checklist	A checklist you tick, then an answer that follows your ticks	2	A judgement call with real trade-offs: you settle what the answer is built on before it is written.
Compare	Add Breakdown	A scored table with a declared winner, and no merged answer	1	Answers that are alternatives to choose between: two designs, two rewrites of one paragraph.
Custom	Add Custom	Whatever your own prompt asks for	1	A fixed output shape none of the three produces: a test plan, a risk table.

Adding a Fuse, Guided or Compare card starts that merge at once. Custom waits, and says so: Just press the Start Merge button when done. Several merge cards can sit over the same answers. Each keeps its own model, its own Retry and its own Remove.

Guided's three stages

Card header	What happens
1/3 · Generating Checklist ...	one request; 3-9 tick boxes come back · - [ ] **Name**: description
2/3 · Criteria Selection ...	no request; tick items, then Confirm Selection
3/3 · Checklist-guided Merge ...	one request; your ticks go back as text

That counter runs in the card header. Stage 1 asks for orthogonal items, especially points of difference. Check All and Uncheck All handle the long lists. Stage 3 sends both what you selected and what you did not. Guided is the only mode that stops and waits for you.

Compare's scored table

| Answer  | Criterion A | Criterion B | Total |
|---------|-------------|-------------|-------|
| reply 1 |    1-100    |    1-100    | 1-100 |
| reply 2 |    1-100    |    1-100    | 1-100 |

Compare picks the criteria itself: up to two on a simple question, up to six on a complex one. Each answer first gets at most one sentence of assessment. Scores are spread across the full 1-100 range. The declaration of a best response comes last, after the table. What that verdict is worth: when the models disagree.

Before a merge can run

Under two answers, on the add card:
  Merge needs 2+ replies. Beam some first.

Starting while cards are still writing:
  Some replies are still generating. Merge what's ready?
  [ Merge now ]   [ Waiting for all... ]

Card header, by state:
  Start Merge      never run
  Retry            has output
  Stop             running
  - Interrupted    after a stop

Two answers with content, and a merge model: that is the whole precondition. If the run finishes while that question is open, the merge starts on its own.

Merging is optional. Shift-click the selected mode and the add card disappears. Any single answer still goes to the chat from its own card (run your first Beam).

The merge model

The Merge Model dropdown in the merge pane sets the model new merge cards get, independent of the models that answered. The model icon on any card retargets that one card. Custom cards each carry their own model, so the dropdown hides while Custom is selected.

Your own merge prompt

Editor field	When you see it
User Instruction:	always editable
System Instruction:	shown with Beam Options -> Detailed Custom Merge

Variable	What it is
{{N}}	the number of answers actually being merged
{{PrevStepOutput}}	the previous stage's output; it is what feeds the Guided merge

Only the Custom card is editable. Your text also runs through Big-AGI's shared prompt variables - {{Today}}, {{LocaleNow}}, the rendering hints. To keep a merge prompt, put it on a persona. Its Custom Merges entries become their own buttons, ahead of the four built-ins free (what a persona is).

Related

• Run several models at once
• When the models disagree
• What a persona is