Everything in this handbook assumes AI belongs in the project. Sometimes it doesn't. The honest answer to "should we use AI here?" is a quiet no more often than the industry pretends. Earn the yes before you commit.
The four-phase workflow in §02 isn't invented. It stands on 20+ years of design frameworks, the methods below shaped how teams discover, define, design, and deliver. AI didn't replace them. It changed how fast each step runs and what one person can do inside them.
Four phases, cyclical. Human at every handoff. The arc hasn't changed in 20 years, what's changed is who runs each step. Agents cover the roles that used to be filled by other people; you cover the judgment that can't be outsourced. Before you touch Phase I, compose your crew.
An agent is an AI that takes multi-step action on your behalf, browsing, running code, reading files, calling APIs, without you at the wheel for every step. Different from a chatbot: you describe the outcome, approve a plan, then review the result. The leverage is real, and so are the failure modes.
skills/*.md.When the off-the-shelf agents don't fit, you can build your own. Same five ingredients regardless of how technical you go: a model, instructions, tools, memory, and an orchestration layer that ties them together.
skills/*.md or equivalent. Iterate here, not in the conversation.Agentic monitoring, also called agent observability, is the production layer that tells you when your agent failed, why, and what to do about it. LLM monitoring watches single calls. Agent monitoring traces entire sessions, tool chains, sub-agent handoffs, and multi-step causal failures. By 2028, Gartner projects 60% of software teams will use AI eval and observability platforms, up from 18% in 2025.
Agents are only as smart as what fits in their context window plus what they can remember across sessions. Memory is where generic AI becomes your AI.
AI flattens some work and sharpens the value of the rest. Specialists go deeper, craft, systems thinking, edge cases. Generalists go broader, AI lets one person credibly span research, copy, visual, and prototype. The best teams pair both, with clear handoffs.
skills/*.md prompts. Generalists inherit craft standards automatically.design-system.md. Generalists propose changes via PR. The design agent reads the canonical file.AI tools are only as good as the context you give them. The same files that guide your team, written as .md, can ground every AI tool in how your brand thinks, writes, and designs. Version-control them next to the work.
brand-voice.md, design-principles.md, and glossary.md as project knowledge. Every new chat starts grounded.design-system.md, your codebase, and design files to build your system. Refine over time..cursorrules or CLAUDE.md. Engineering and design pull from the same source of truth.skills/*.md files in a shared repo. Anyone can load them into Claude or ChatGPT for consistent team workflows.When you send a single message, 11 distinct steps run in sequence, and usually multiple times per turn as the agent loops through tools. Each step has its own data shape, its own latency, and its own failure mode. Understanding them explains why latency spikes, why context fills up, and why tool-heavy turns are expensive.
~300–800ms latency./v1/messages. Streaming responses use SSE (server-sent events).429.~25 tool calls per turn.system prompt → tools → memories → retrieved context → conversation → current turn. Later content gets stronger attention.1M tokens for Claude Sonnet 5.5, 400K–2M for GPT-6 and Gemini 3.6. Even with huge windows, attention quality degrades past ~200K, chunking and retrieval still matter.~50K–200K tokens.$3–$15/MTok input, $15–$75/MTok output, a 200K-token agent run with 10K output easily hits $1–$3. Prompt caching cuts repeat-context cost by up to 90%.~2,048–12,288 floating-point numbers. This is the model's dimensionality, higher = more expressive, more expensive.[sequence_length × hidden_dim] tensor. For a 1,000-token prompt at 8,192 dims, that's 8.2M numbers before any reasoning has happened.~40–120 layers deep. Each layer: multi-head self-attention → feedforward (MLP) → residual + layer norm.O(n²) cost in sequence length, why long contexts get quadratically slower without tricks like flash attention, sliding windows, or state-space layers.~128K scores.~50–200 tokens/sec for frontier models in 2026, up from 20–100 a year ago. Distilled and smaller models hit 500+ tokens/sec. Why you see text appear word-by-word.200ms–2s. Dominated by context size, a 100K-token prompt takes longer to "prefill" than a 1K one.50–200 tokens/sec of streaming output for 2026 frontier models. Smaller / distilled models hit 500+ tokens/sec. Reasoning models (Claude Extended Thinking, GPT-6 reasoning, Gemini 3.6 Thinking) can pause for seconds-to-minutes before any output.200K–2M tokens. But attention isn't uniform, models often weight the start and end of context more than the middle ("lost in the middle").5–10× cheaper than output tokens. Tool-heavy turns that fold tool output back into context can silently balloon costs.90% cost reduction for repeated system prompts or long docs.temperature=0, outputs can vary between runs due to floating-point non-associativity on parallel GPUs. Reproducibility requires seeds and identical hardware.~1,500 tokens each; a 1-minute video can run ~15K tokens. Design for that math.1M–10M tokens across 10–100 tool calls. Observability is required to see what actually happened.A world model is an AI that learns how physical environments behave, gravity, momentum, spatial consistency, cause and effect, and can simulate or generate them. Not a language model, not an agent. An entirely different class of system, and an entirely different design surface. If you work on video, 3D, spatial interfaces, robotics, AR/VR, or game-like experiences, this is where the ground is shifting fastest in 2026.
Four concrete examples of designer-builders using an agentic team to compress what used to take weeks into days. Each shows the chain of agents, the human checkpoints, and the tools behind each step. Not aspirational, these are shipping today.
brand-voice.md you can version alongside the project.AI security is a different animal from traditional app security. Your agents read untrusted content, call tools you authorized, and persist memory you may have forgotten about. The usual perimeter model doesn't hold, the attack surface is inside the context window. Treat this as table stakes, not optional hardening.
AI bias is a design problem, not just a model problem. Every frontier model inherits bias from its training data, the open web, licensed text, and whatever human preferences were baked in during fine-tuning. By the time an agent runs in your product, that bias has been amplified by your prompts, your tool choices, your UX, and the user's own context. You own the bias your agent exhibits in production, regardless of who trained the model.
skills/*.md teach the agent what "good" looks like. If your examples skew one direction, so will your output. Audit your own skills periodically.The vocabulary of AI moves fast. These are the terms used throughout the handbook, written for designer-builders, plain language, no PhD required.
Enter your email to unlock all 11 chapters — frameworks, agents, memory systems, security, and more.