Skip to content

Feedback Fragments

A feedback fragment is a kind of Interactive Fragment that, instead of steering the Writer, produces a short out-of-character note for you after the reply is written.

The inverted direction

Ordinary interactive fragments point inward: the Director fills them in to shape what the Writer produces (AI → AI). Feedback fragments point outward: once the reply is finished, the model steps out of character and writes a note to the player (AI → user). The note is shown in the Inspector panel and never reaches the Writer or affects the story.

Think of it as a game master leaning over to give you a tip between turns.

What it's good for

Anything you want the model to tell you rather than weave into the prose:

  • Suggestions — "2 fresh things you could do next" (this ships as a built-in, disabled by default).
  • Coaching — pacing notes, reminders of dangling threads, tone observations.
  • Meta status — anything that would break immersion if it appeared in-scene.

How it differs from other fragments

Feedback fragments are authored exactly like any other interactive fragment — same ID, Label, Description, Required flag, and ordering — but with the field type set to feedback (note to you). The differences:

  • They run in a separate post-writer step, after the reply (and any editor edits) is final, so the note can react to the actual finished text.
  • Their Injection label is used as the heading for the note in the Inspector.
  • They're collected together into a single give_feedback call — you can have several, and each becomes one row in the feedback card.

Enabling it

Feedback fragments are gated behind the Editor Feedback feature flag in Settings → Agents. Until you turn that on, any feedback-type fragments are greyed out in the fragment list and won't run. Enable the flag, then enable the individual feedback fragments you want.

No extra cost on the prompt

The feedback step reuses the same cached prompt the Writer and Editor already built and simply forces the give_feedback tool at the end, so it doesn't pay a fresh prompt-processing cost. See KV Cache Reuse for the caching design.