User Story:

As a form creator (Maya), in order to control how applicants experience the form, I want to shape the form structure using natural language and manual controls, with a live preview of my changes

Preconditions:

• A FormProject exists with extracted DataCollectionSpec and FormSpec (Slice 2)
• Maya is authenticated

Acceptance Criteria:

• Maya can view the current FormSpec for a project
• Maya can describe changes in natural language (e.g., “add an eligibility screener before page 5”)
• LLM proposes concrete FormSpec edits based on Maya’s intent and the underlying DataCollectionSpec
• Maya sees a clear before/after of proposed changes
• Maya can accept or reject proposed changes
• Maya can manually fine-tune: reorder pages, move groups between pages, set delivery mode per section
• LLM suggests delivery modes based on section complexity
• Live preview shows what Carlos would see
• Changes are saved to the FormSpec

Success Metrics:

• Maya can complete a form shaping session in under 10 minutes
• LLM-assisted edits align with Maya’s stated intent
• LLM delivery mode suggestions align with section complexity

Notes:

• LLM-assisted authoring is the primary workflow; manual controls are for fine-tuning
• Story 5 handles proposal/draft semantics — Story 4 edits FormSpec directly
• Live preview renders the form as Carlos would see it
• FormSpec changes are persisted via existing ProjectStore

Definition of Done:

• Acceptance criteria met
• Threat model updated – any new trust boundaries, data flows, or attack surfaces are reflected in catalog/architecture/threat-model.md
• Technical documentation updated – architecture docs and decisions are current
• Tests pass
• Type checking passes
• CI pipeline green
• Deployed and demoable