targetable-feedbacklisted

# Targetable feedback ## When to activate Two paths, both keyed on user intent. **Generative.** User asks for a draft and signals iteration intent in the same request. Generate the content and render it directly as the widget. No plain-prose version first. > "Draft an RFC for the new auth system and let me give feedback on each section." **Retroactive.** User wants to mark up multi-section output already in the conversation. The more common path. > "Let me give feedback on each part." / "Make that targetable." / "I want to iterate on what you just wrote." **Ambiguity boundary.** Single-edit requests are not activations. "Fix bullet 3" is one-shot feedback, answer inline. "Let me refine each bullet" is iteration intent, activate. If the user names specific edits in the activation message, apply those edits first, then render the result. ## The verbatim rule (CRITICAL) When activating retroactively, the prior assistant content goes into the widget unchanged. Do not rewrite, summarize, expand, sharpen, or improve units during conversion. The user wants to mark up what they saw, not negotiate a new version. Parsing is deterministic. A bullet reading "Goal 1: reduce time" becomes a unit reading "Goal 1: reduce time," not "First goal: time reduction." **Do not re-segment the input.** Paragraph boundaries (blank lines) and bullet boundaries are inherited directly from the source. A paragraph that contains multiple sentences becomes ONE unit, even if the sentences address