# AI for Product Management & Design URL: https://justuju.org/role/product/ For your role AI changes product work twice. It speeds up your own craft, from research synthesis to working prototypes, and it becomes a material you design with, one that behaves differently every time. ## How AI is changing this work - Prototypes are built from a prompt in an afternoon, so ideas are tested with something real. - Research synthesis across hundreds of interviews and tickets takes hours, with quotes linked to sources. - AI features need a new skill: defining good output and measuring it, because the same input can give different results. ## Key layers, most important first ### 1. Layer 9 · Applications & Interfaces URL: https://justuju.org/layer/applications-interfaces/ Why it matters: This is your own layer: the interface where people meet AI. Chat, voice, a button inside a page or an assistant beside the work are design choices with different costs. What to do here: - Choose the form: A chat box is not the default. Often a button, a suggestion or a filled-in field serves better. (see https://justuju.org/concept/embedded-ai/) - Design for being wrong: Show sources, make edits easy and give people a way back, because some outputs will be wrong. (see https://justuju.org/concept/copilots/) - Synthesis with receipts: Cluster interviews and feedback into themes, each linked to the clips and quotes behind it. (see https://justuju.org/concept/chat-assistants/) Examples as of 2026-11-11: Figma Make (https://www.figma.com/solutions/prompt-to-app/); Dovetail (https://dovetail.com/product/ai-projects/); Notion Agent (https://www.notion.com/help/notion-agent) ### 2. Layer 8 · Agents URL: https://justuju.org/layer/agents/ Why it matters: Agents that write code turn a described idea into a working app. They let you test real behaviour before engineering time is spent, and they are a feature type you may ship. What to do here: - Prompt to prototype: Describe the app, refine by chatting, share a live link with five users this week. (see https://justuju.org/concept/coding-agents/) - Plan before build: Ask the agent for a plan first, correct it, then let it write. (see https://justuju.org/concept/agent-loop/) - Agent features need limits: If your product acts for users, decide what it may do alone and what needs a tap to confirm. (see https://justuju.org/concept/agent-permissions/) Examples as of 2026-11-11: Lovable (https://lovable.dev/build/app-builder); v0 by Vercel (https://v0.app/); Cursor (https://cursor.com/docs/agent/overview) ### 3. Layer 7 · Orchestration & Tooling URL: https://justuju.org/layer/orchestration-tooling/ Why it matters: For an AI feature, this layer is the product: the instructions, the data it is given and the tests it must pass. Product owns the definition of good. What to do here: - Evals as the spec: Write 50 to 100 example inputs with what a good answer looks like; that set is your requirements. (see https://justuju.org/concept/guardrails-evals/) - Context is the lever: Most quality gains come from what you put in front of the model, not from a bigger model. (see https://justuju.org/concept/context-engineering/) - Ground it: Give the feature your real content to answer from, and show the source. (see https://justuju.org/concept/rag/) Examples as of 2026-11-11: LangSmith (LangChain) (https://www.langchain.com/langsmith/evaluation); Braintrust (https://www.braintrust.dev/); promptfoo (now part of OpenAI) (https://openai.com/index/openai-to-acquire-promptfoo/) ### 4. Layer 6 · Inference & Optimization URL: https://justuju.org/layer/inference-optimization/ Why it matters: Every AI feature has a price and a wait per use. Both are set here, and both belong in the product decision. What to do here: - Price per action: Estimate cost per use from input and output length before you design the pricing page. (see https://justuju.org/concept/model-apis/) - Right-size the model: Use a small fast model for simple steps and a large one only where quality demands it. (see https://justuju.org/concept/model-apis/) - Cache what repeats: Reusing the long, unchanging part of a request cuts cost and wait. (see https://justuju.org/concept/batching-caching/) Examples as of 2026-11-11: OpenRouter (https://openrouter.ai/docs/quickstart); Amazon Bedrock (https://aws.amazon.com/bedrock/); Claude API pricing (https://platform.claude.com/docs/en/about-claude/pricing) ## High-leverage use cases - Clickable prototypes: Working prototypes from a prompt or an existing design, for user tests the same week. - Research synthesis: Themes, quotes and clips from interviews, tickets and reviews, traceable to the source. - Specs and first drafts: Requirement docs, release notes and user stories drafted from notes, then edited. - Eval sets for AI features: A maintained set of examples that defines good output and runs on every change. - Cost and speed modelling: A simple model of cost and wait per user action, before committing to a design. ## Start here: the next 30 to 60 days - Week 1 to 2: Build one prototype with a prompt-to-app tool and put it in front of five users. - Week 3 to 4: Run your last ten interviews through an AI synthesis tool and compare with your own notes. - Day 30 to 60: For one AI feature, write 50 examples of good and bad output with engineering, and run them on every change. - Day 30 to 60: Work out the cost and wait per use for that feature at your expected volume. ## Risks and pitfalls - Prototype mistaken for product: Generated apps demo well and hide gaps in security, data handling and edge cases. Label prototypes as prototypes. - Shipping without evals: If you cannot measure quality, every model or prompt change is a gamble. Build the test set first. - Synthetic users: A model's guess at what customers think is not research. Use it to prepare interviews, not to replace them. - Costs that scale with success: AI features cost money per use. A popular feature with no cost model can erase its own margin. - Features that act on untrusted text: If your feature reads web pages, emails or uploads, hidden instructions in them can steer it. Source: https://genai.owasp.org/llm-top-10/ ## Sources - Figma Make (figma.com): https://www.figma.com/solutions/prompt-to-app/ - Dovetail (dovetail.com): https://dovetail.com/product/ai-projects/ - Notion Agent (notion.com): https://www.notion.com/help/notion-agent - Lovable (lovable.dev): https://lovable.dev/build/app-builder - v0 by Vercel (v0.app): https://v0.app/ - Cursor (cursor.com): https://cursor.com/docs/agent/overview - LangSmith (LangChain) (langchain.com): https://www.langchain.com/langsmith/evaluation - Braintrust (braintrust.dev): https://www.braintrust.dev/ - promptfoo (now part of OpenAI) (openai.com): https://openai.com/index/openai-to-acquire-promptfoo/ - OpenRouter (openrouter.ai): https://openrouter.ai/docs/quickstart - Amazon Bedrock (aws.amazon.com): https://aws.amazon.com/bedrock/ - Claude API pricing (platform.claude.com): https://platform.claude.com/docs/en/about-claude/pricing - OWASP Top 10 for LLM Applications: https://genai.owasp.org/llm-top-10/ - NIST AI Risk Management Framework (AI RMF 1.0): https://www.nist.gov/itl/ai-risk-management-framework --- Justuju · https://justuju.org/ · Content CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Quote with a link to the page.