A practical walkthrough

AI for Product.

Tools, concepts, and pitfalls — from one prompt to a working deliverable.

AuthorJian Zhao

Chapters6

Reading time20 minutes

Date2026.04

Table of Contents

One prompt, one deliverable.

A single structured prompt that produces a working deck in a few minutes.

prompt.md · copy & paste

You are a product designer. Use the frontend-design skill (~/.agents/skills/frontend-design) for this task.

I'm giving a 20-minute talk on "AI for Product" to our team. Please build a single index.html deck with keyboard navigation that contains:

Cover — title, author, reading time
Table of contents — clickable, jumps to each chapter
Five chapters — a warm-up demo, an AI tools overview, basic concepts, everyday deliverables, and common pitfalls
Q & A closer

Style: light theme, understated tech feel, 12px radius, single blue accent. Include a dark mode toggle.

Support English, Chinese, and Korean with a language switcher.

Paste into your AI tool and press return — no coding required.

02 Pick Your Tool

Start with the starred ones.

Nine tools, six use cases. The three starred cards are the ones we'd install first. Hover any card for details.

01Assistant

Claude — Desktop · Code

02Assistant

Gemini — Web · CLI

03Search

Perplexity

04Research

NotebookLM

05IDE

Cursor

06IDE

Windsurf

07Coding

Codex — Desktop · CLI

08Design

Pencil

09Writing

Markdown — Typora · Obsidian

03 Basic Concepts

Skill, protocol, tools.

The shared vocabulary for the rest of this deck — what they mean and where the AI gets its superpowers.

I.Capability

Agent Skills

Reusable recipes the AI follows step-by-step. In practice: ready-made playbooks that produce PRDs, slides, or research summaries in a consistent shape.

~/.agents/skills/frontend-design
~/.agents/skills/canvas-design

Browse & install at skillsmp.com →

II.Protocol

MCP

A standard way to connect AI to your tools. In practice: it's how the AI can read your Figma, Linear tickets, or Notion docs directly.

mcp: filesystem · github · figma · slack · postgres · playwright · …

modelcontextprotocol.io →

III.Shell

CLI Tools

Command-line utilities the AI can run on your behalf. In practice: the AI pulls real data (GitHub issues, logs, files) instead of making things up.

$ gh pr view # GitHub
$ rg "TODO"   # search
$ jq '.data'   # JSON
$ ffmpeg …   # media

Example: GitHub CLI →

04 Daily Deliverables

Three headlines, five extras.

Agent Skills turn prompts into actual .pptx / .docx / .html files — not just text for you to copy out.

pptx Slides

Produce a full .pptx deck from an outline — talking points, visuals, and speaker notes already in place.

›Turn this outline into a 10-slide Q2 review deck.

docx Docs

Write PRDs, memos, and proposals as properly formatted Word files — tables of contents, headings, and styles included.

›Draft a PRD for the onboarding redesign as a .docx.

frontend-design Pages

Product intro pages, launch landing pages, interactive demos — no code required.

›Build a landing page for our new AI meeting tool.

Five more worth knowing

xlsx Competitive tables, data cleanup, dashboards.

canvas-design Posters, launch images, internal comms art.

pdf Merge, split, fill, extract — PDFs handled.

doc-coauthoring Structured collaboration for PRDs & spec docs.

internal-comms Status updates, FAQs, incident reports.

Install Skills live in ~/.agents/skills/. Browse and one-click install the full library:

skillsmp.com → github →

05 Common Pitfalls

Six mistakes to avoid.

The issues people hit most often — and how to side-step each one.

01Problem

Hallucination

Confident, fluent, and wrong. The model invents facts, numbers, or citations when it doesn't know.

Fix Ask for sources. Cross-check numbers against the original doc. Give it your data instead of asking it to recall.

02Problem

Context Loss

Long chats run out of memory; the model gradually drifts from what you originally asked.

Fix Restate the goal every ~10 turns. Start a new chat when it stalls. Keep a reusable context file you can paste back in.

03Problem

Over-Trust

The first answer looks polished, so you ship it. But polished ≠ correct, and the model often just agrees.

Fix Ask for 3 options, or "argue against this." Never ship the first draft without a 60-second read-through.

04Problem

Flattened Insights

AI summaries of interviews, reviews, or feedback smooth over outliers and emotion — the very signal you need.

Fix Always read a few raw transcripts. Ask for quotes, not summaries. Request "3 contradicting views" instead of themes.

05Problem

Stale Knowledge

Training data has a cutoff date. Competitor features, pricing, or tool updates may be silently out of date.

Fix Turn on web search for anything current. Give it the latest link. End with: "say 'not sure' if you don't know."

06Problem

Data Leakage

Pasting roadmaps, customer quotes, or revenue into consumer chat tools can expose unreleased strategy.

Fix Use the enterprise/team tier for work data. Disable training on your account. Strip customer names before pasting.

06 Q & A

Now try it.

Your questions, please.