A one-page operating model for Claude Code architecture and CCA-F certification prep — synthesized from the last five LearnTeachMaster.org articles.

REFERENCE SHEET

The Governed Agentic Architecture

A one-page operating model for Claude Code architecture and CCA-F certification prep — synthesized from the last five LearnTeachMaster.org articles.

The Core Idea

The model is not the architecture. Claude Code can reason, inspect a repository, write code, and delegate work — but capability without structure leaves engineering value on the table. Two things solve this, and they are not the same thing:

• Project structure tells the agent where it is.

• Intent tells the agent where it needs to go.

Every layer below exists to answer one of those two questions, plus a third: what proves the work is actually done.

The Architecture Stack

CLAUDE.md

Where am I? — repository purpose, architecture, stack, conventions, key commands. Keep small: a map, not the contents of every building on it.

INTENT.md

Where am I going? — scoped goal, inputs/context, constraints, required outputs, success criteria for this task.

Rules

How do we behave here? — scoped conventions (API, testing, security, DB) that don't belong in permanent context.

Skills

How do we do this, organizationally? — reusable procedural knowledge, loaded when needed, not held permanently in context.

Agents

Who reasons about a bounded piece? — used only for isolation, specialization, parallelism, or independent review. Not a default.

Tools / MCP

What can the system reach? — governed access to repos, APIs, databases, ticketing, cloud, CI/CD.

Hooks

What cannot be optional? — deterministic enforcement: tests, security scans, formatting, prohibited actions.

Settings

What is the system permitted to do? — the permissions boundary, independent of what it knows or reasons.

Validation / Evidence

How do we prove it worked? — tests, scans, schema checks, traceability. Generation is not completion.

Two Distinctions Worth Memorizing

Skills vs. Agents

Skills teach the AI how we do something. Agents reason about a bounded problem. A large percentage of what looks like an “agent” requirement is actually just organizational knowledge that belongs in a Skill. Default to Skills. Reach for an Agent only when you need isolation, specialization, parallelism, or independent review — not because the capability exists.

Intent vs. Context

Context engineering determines what the system needs to know. Intent-driven engineering determines what the system needs to accomplish. Perfect context without precise intent still leaves the outcome undefined. Both are required; neither substitutes for the other.

The Execution Loop

Understand  →  Plan  →  Execute  →  Observe  →  Validate  →  Correct  →  Prove

The loop does not end at “generate.” It ends when success criteria are demonstrably met, or a defined stop condition escalates to a human. Generation is easy; proof is the hard part — a beautiful diagram, a professional-looking Terraform file, or a confident ADR can all be wrong.

Design-Time Checklist

Run this before delegating any non-trivial task to Claude Code:

☐  Repo map exists (CLAUDE.md-equivalent) and stays small

☐  Minimum sufficient context defined at every boundary crossed

☐  Scoped intent artifact defines goal, inputs, constraints, outputs, success criteria

☐  Non-negotiables are enforced by Hooks, not by instruction alone

☐  Required procedural knowledge exists as a Skill, not re-explained in-prompt

☐  Permissions boundary (Settings) matches what the system should be allowed to do

☐  Agent use is justified by isolation, specialization, parallelism, or review — not default

☐  Evidence of success is defined before work starts, not inferred after

The Anti-Pattern to Avoid

Don't build an agent zoo. A repository does not become more agentic because it has twelve agents, twenty skills, and fifty hooks. Every agent is a context boundary; every boundary is a failure path. The rule that eliminates most unnecessary complexity:

Use the minimum sufficient architecture required to satisfy the intent safely and reliably.

Sometimes that's Intent → Claude → Tests → Done. That is a complete, correct architecture when the problem doesn't require more.

CCA-F Mapping

This stack maps directly onto certification material — it's one architecture viewed from two angles, not two separate things to study:

• Skills, Hooks, MCP, Subagents mechanics → the capability and enforcement layers above

• Context engineering → what crosses each boundary, and why less is usually better

• Governance and permissions → Settings + Hooks together define how much autonomy is safe

• Validation and evidence → the exam's emphasis on proof over generation

Source Articles

From Developer to AI Architect: The Autonomous Architectural Workbench — learnteachmaster.org, Aug 21, 2026

Claude Code Gets Better When Your Project Gives It Structure — learnteachmaster.org, Aug 21, 2026

Intent-Driven Engineering Meets Graph Engineering — learnteachmaster.org

Claude Code Plugins at Scale: Building an Enterprise Shared Services Strategy — learnteachmaster.org

TeamBrain: From Sales Handoff Chaos to a Living Project Brain — learnteachmaster.org, Dec 16, 2025

Mark Kendall  ·  Intent-Driven Engineering  ·  Page