# UX Direction: Agentic Steering in Transitions

## The Problem
Linear chat is a poor interface for structural editing. When an AI proposes a multi-point plan or idea in a chat bubble, forcing the user to type "Change point 2 to X and make point 4 about Y" creates a high cognitive load. It treats the AI like a generic chatbot rather than a collaborative agent.

## The Goal
Provide users with a way to steer the AI's generation process immediately before a major context shift (like moving from Chat ideation to Outline generation), without forcing them to type complex prompt-engineering instructions.

## Phase 1: The Steered Transition (Implemented)
Right now, transition buttons (like "Create Outline Now") act as binary triggers. By adding an optional textarea directly into the Call-To-Action (CTA) block, we change this from a binary trigger into a **steered transition**.

**Flow:**
1. AI proposes an idea in chat.
2. The CTA block appears:
   - **Textarea:** "Any specific focus or tweaks before I build the outline? (Optional)"
   - **Button:** "Create Outline Now"
3. When clicked, the user's feedback is injected directly into the outline generation prompt.

This captures global steering ("Make sure it sounds professional" or "Skip the history section") with extremely low friction.

## Phase 2: The Structured Canvas (Future Consideration)
Instead of trying to hack inline comments into static markdown chat bubbles, granular editing should happen where structure already exists: **The Outline View**.

**Proposed Flow:**
1. The user accepts the rough idea in Chat and clicks "Create Outline Now".
2. The UI flips to the Planning Tab, rendering the JSON outline as editable cards.
3. Every section card gets an **"🪄 AI Refine"** button/input.
4. The user can type localized feedback on a specific card: *"Make this section focus more on X"*.
5. The AI regenerates *just that specific JSON section*.

This completely separates ideation (Chat) from structural editing (Outline UI), giving the user "Antigravity-style" inline commenting where it belongs: on the actual structured data.