Generative Features

Context

Your team shipped an AI text generator into the product. The demo was impressive. Now the support tickets roll in: “The output is wrong”, “This doesn’t sound like me”, “Why can’t I just tweak this part?”

The problem isn’t the model. The problem is that generative features require a fundamentally different UX approach than deterministic features. A button always delivers the same result. A generative feature delivers something different every time — and your user has to decide whether it’s good enough.

Concept

The Core Challenge: Non-Deterministic Output

Generative features have a unique challenge: quality varies. Every output is a draft, not a final result. This creates three tension points:

Evaluation burden: The user must assess every output themselves — that costs cognitive energy
Expectation gap: AI is probabilistic, stakeholders expect deterministic. Figma 2025: 78% say AI boosts efficiency, but under 50% say it makes them better
Draft-to-final gap: Between “first draft” and “finished output” lies a workflow you need to design

UX Patterns by Modality

Text Generation:

Inline generation (Notion AI): Slash command, output appears directly in the document. Lowest friction, but limited control
Side panel (ChatGPT Canvas, Claude Artifacts): Separate panel with direct editing. More context, more control
Highlight-and-instruct (Canvas): Select text, request targeted changes. Most precise method
Visual diff (Notion): Deletions in gray, additions in blue. User sees immediately what changed

Image Generation (Midjourney paradigm):

4-image grid, select, variations (subtle/strong), upscale, refine, export
Never force commitment to a single generation
Vary Region for partial regeneration — change only the part that doesn’t work

Code Generation:

Inline completion (Copilot): Ghost text, Tab to accept. Minimal interruption
Prompt-to-code (v0): Chat to complete components, live preview, iterate

Edit/Refine Workflows

The output is rarely perfect on the first try. Your product needs at least two of these mechanisms:

Conversational refinement: “Make it shorter” — natural language iteration
Direct editing: User types directly in the AI output panel
Variation branching: Generate multiple alternatives from a single base
Behavioral contracts: Define what the AI can do, can’t do, and should ask about before acting

Quality Guardrails

Input guardrails: Filter before the model sees the prompt
Output guardrails: Check results before the user sees them
Fallback UX: When output gets blocked — don’t just show an error, offer an alternative
False positive tension: Legitimate prompts get blocked. Too strict = frustrating, too loose = risky
Latency impact: Every guardrail costs time. Communicate wait time, not just the result

Framework

Generative Feature Design Matrix:

Dimension	Question	Low Maturity	High Maturity
Output control	Can the user modify the output?	Regenerate button only	Edit, refine, variations
Expectation mgmt	Does the user know what to expect?	No guidance	Onboarding, examples, limitations
Guardrails	How is quality ensured?	Output filter only	Input + output + fallback UX
Iteration	How does the user refine?	Copy-paste into new prompt	Conversational refinement + direct edit
Feedback loop	Does the system learn from the user?	No feedback	Thumbs, edits as signal, behavioral contracts

Minimum Viable Generative Feature: Output control + expectation management + at least one refine mechanism. Without these three, you’re sending users into a dead end.

Scenario

You’re a PM at a content marketing SaaS. Your new feature: AI-generated blog drafts from briefings. Beta data after 4 weeks:

2,400 drafts generated, 68% get further edited (good)
Average 4.2 regenerations per draft before users are satisfied (bad)
31% of users give up after 2+ regenerations and write from scratch
Feature NPS: +12 (core product: +38)
Top complaints: “Sounds generic”, “Can’t adjust tone”, “Changes always affect the entire text”

Options:

Model upgrade: Better base model, hope for better first-draft quality. Cost: $40k/year more in API costs
Refine workflow: Highlight-and-instruct + tone settings + visual diff. Cost: 6 weeks engineering
Variation-first: 3 variants per generation + mix-and-match function. Cost: 4 weeks engineering, 2x API costs

Decide

How would you decide?

The best decision: Option 2 — Refine workflow.

Why:

The problem isn’t first-draft quality (68% edit rate is solid) — the problem is users can’t make targeted adjustments to the output
4.2 regenerations per draft shows: users know what they want, but lack the tools to communicate it to the system
Highlight-and-instruct directly solves the top complaint “changes always affect the entire text”
Tone settings as a behavioral contract: users define upfront what the AI should do

Why not the others:

Option 1: A better model might reduce regenerations from 4.2 to 3 — but without an edit workflow, users stay stuck in the regeneration loop
Option 3: Variations help with selection, not refinement. The core pain is “I can’t change anything specifically”

Reflect

Generative features are workflows, not outputs. The generate button is the beginning, not the end. Without edit/refine mechanisms, you’re delivering a slot machine instead of a tool.
Expectation management is product design. If you don’t communicate that AI works probabilistically, you create disappointment. Onboarding honesty (“I’m still learning”) isn’t admitting weakness — it’s UX.
Guardrails cost latency, no guardrails cost trust. Find the balance and communicate wait times transparently.
User segments differ radically. Embracers want power features, skeptics need safety nets. Segment by AI attitude, not just by role.

Sources: Figma AI Report (2025), Notion AI UX Patterns (2024), Midjourney UX Documentation (2024), Lenny’s Newsletter — User Segmentation for AI (2024), OpenAI Canvas Design Blog (2024)