Noddle Deck

Agent Engineering

Skills, slash commands, hooks, MCP, harnesses, and the multi-agent systems built on top of them — explained from first principles, by the Noddle Deck team.

Agent Engineering

Agent Skills, Explained: Teach Your AI Agent Like You'd Onboard an Engineer

Skills package expertise so an agent can load it only when the task calls for it. Here's the anatomy, the lifecycle, and the mistakes that make skills invisible to the agent that should be using them.

15 min read
agent skillsclaude codecontext engineering
Read post

Agent Engineering

Slash Commands: Reusable Prompts You Can Ship

A slash command turns a prompt you'd otherwise retype into a versioned, shareable action — the difference between a good prompt and a piece of team tooling.

14 min read
slash commandspromptsdeveloper workflow
Read post

Agent Engineering

Hooks: Deterministic Guardrails for Non-Deterministic Agents

Hooks run outside the model's judgment entirely — deterministic checks that fire before or after a tool call so the same rule holds every single time, not just most of the time.

14 min read
hooksguardrailsautomation
Read post

Agent Engineering

MCP: The USB Port for AI Agents

The Model Context Protocol gives agents one standard way to talk to any tool or data source, instead of a bespoke integration for every model-and-service pair.

15 min read
mcpintegrationstool use
Read post

Agent Engineering

The Agent Harness: Everything Around the Model

The model is one component. The harness — the loop, the tool router, the context manager, the permission system — is what actually turns it into a working agent.

12 min read
agent harnessarchitecturetooling
Read post

Agent Engineering

Multi-Agent Systems: Orchestrating a Team of AIs

Splitting a task across a lead agent and several subagents can beat one big context window — if you get the handoffs and the shared state right.

11 min read
multi-agentorchestrationarchitecture
Read post

Agent Engineering

Loop Engineering: Agents That Run While You Sleep

Long-running agent loops need their own discipline — checkpointing, budget limits, and exit conditions — or they either stall out or burn through your quota unattended.

11 min read
loop engineeringautomationagents
Read post

Agent Engineering

Inside the Claude Opus System Prompt: an Architectural Teardown

A leaked 183 KB claude.ai system prompt, measured section by section: what's actually behavior versus plumbing, where the identity sentence sits, and what three model releases patched.

16 min read
system promptsclaudeprompt engineering
Read post

Agent Engineering

The Claude Sonnet System Prompt as a Control Loop

Sonnet's behavior isn't fixed by a document loaded once at the top of a chat. It's maintained turn by turn — a static base, per-turn injections, classifier-triggered payloads, and drift reminders that fire only once a conversation has run long enough to need them.

14 min read
system promptsclaudeprompt engineering
Read post

Agent Engineering

The Claude Fable 5 System Prompt: What a New Model Tier Looks Like on Paper

A leaked prompt repo shows what actually changes on paper when a vendor ships a new model tier — and the most interesting finding isn't what got added, it's what got shorter.

17 min read
system promptsclaudemodel tiers
Read post

Agent Engineering

What an AI Architect Actually Does

Someone ends up owning model selection, the harness, the skill library, evals, cost budgets, and guardrails as a single system — whether or not their org ever writes the title down.

21 min read
ai architectagent engineeringcareer
Read post

Agent Engineering

How to Build an AI Data Platform Your Business Applications Can Search

Most 'AI search' is a chatbot wired straight to a database, and it fails for a predictable reason: the model has nothing trustworthy to stand on. Here's the reference architecture that builds discovery before it builds the agent — and the adapter pattern that keeps business apps decoupled from it.

16 min read
ai platformsemantic layeragent engineering
Read post

Build in Public

One canvas, three tools: how a crew of three ships Noddle

The Noddle family in plain terms — Board, draw, Deck, and Docs next. Why three people ship one product at a time, and where each tool honestly stands.

4 min read
build-in-publicnoddleannouncement
Read post

Data Engineering

Data Reconciliation, Explained: Proving Your Data Survived the Move

A successful copy job is not evidence. Here's how reconciliation actually works when data moves from zone A to zone B — the four levels of checks, the principles that keep comparisons honest, and the metrics that make it governable.

19 min read
data reconciliationdata governancedata engineering
Read post