You have built each piece separately. CLAUDE.md in Lesson 2. Skills in Lesson 3. Hooks and memory from earlier chapters. Now let's see the full picture — everything working together as one system.
The five layers
Your agent setup is not a collection of config files. It is a system with five layers, each serving a distinct role.
CLAUDE.md
Permanent project rules. Loaded every session. Under 200 lines. The foundation that never changes between tasks.
Skills
On-demand knowledge. Loaded when triggered. Zero cost until invoked. Deep workflows for specific tasks.
Hooks
Deterministic enforcement. 100% reliable. Zero context cost. Things that must always happen, without exception.
Memory
Evolving personal and project context. Loaded when relevant. Your preferences, corrections, and accumulated knowledge.
Verification
Automated quality gates. Catches what you would miss. The safety net that makes speed safe.
How they interact
Each layer has a different job, but they work as a unit. Here is how they connect.
CLAUDE.md tells Claude the rules — where files go, what can import what, what to never do. This is the baseline. Every session starts here.
Skills provide deep knowledge for specific tasks. When you invoke /deploy, Claude gets the full deployment checklist. When you invoke /review, Claude gets the review criteria. Skills add depth where CLAUDE.md provides breadth.
Hooks enforce non-negotiable quality. CLAUDE.md says "run lint after edits." A hook makes sure it actually happens, every time, regardless of context pressure. Hooks are the guarantee that rules are not just suggestions.
Memory remembers your preferences across sessions. CLAUDE.md says what the project needs. Memory says what you prefer. The two together give Claude both the project rules and the personal context.
Verification catches what slips through. Even with rules, skills, hooks, and memory, mistakes happen. Verification is the final checkpoint — build, types, lint, tests, security, diff review.
A real task through the full system
Let me walk you through what happens when I say "add a contact form" with my complete setup active.
CLAUDE.md kicks in first. Claude reads the routing table and knows: a new feature component goes in apps/web/components/. It checks the boundary rules — the component needs TRPC for data, so it will use @jd/api (not a REST route). It sees the anti-patterns — never import @jd/shared/server in a client component.
Memory loads next. Claude sees I prefer functional components with TypeScript. It knows I use Tailwind for styling. It remembers I like form validation with Zod schemas.
I invoke /feature. The skill loads and Claude follows the workflow: confirm the spec, create a branch, plan the implementation, build in small steps. Each step is guided by the skill.
Hooks fire automatically. Every time Claude edits a file, the PostToolUse lint hook runs ESLint. If Claude tries to write to .env, the PreToolUse hook blocks it. These fire silently, in the background, without consuming context.
Before commit, I run /verify. The 6-phase pipeline runs: build passes, types check, lint is clean, no secrets in the diff. All green.
The commit skill handles the rest. Stage the right files, generate a message, create the commit.
The entire flow — from spec to commit — happened with minimal intervention from me. I confirmed the spec, reviewed the code, and approved the commit. The system handled the rest.
The ECC harness scoring
The ECC framework formalizes this with a 0-70 point scoring system. Seven categories, each scored 0-10.
Category Score
─────────────────────────────────
CLAUDE.md quality /10 Structure, specificity, boundaries
Agent definitions /10 Custom agents, role specialization
Hook coverage /10 Lifecycle events, reliability
Skill inventory /10 Coverage of major workflows
MCP configuration /10 Connected tools, relevance
Security posture /10 Permissions, env protection, secrets
Cost optimization /10 Context budget, lazy loading, scoping
─────────────────────────────────
Total /70
You do not need to score yourself on all seven right now. The four-dimension audit from Lesson 1 gives you the macro view. The ECC categories give you the micro view — useful when you want to drill into a specific area and improve it.
My current blueprint
Here is my actual setup. Not aspirational — current. What I use daily on production projects.
CLAUDE.md: 150 lines. Commands, structure, boundaries, anti-patterns, routing table. No reference information. No documentation. Pure rules.
Skills: 8 total.
/deploy— pre-deployment checklist/review— code review workflow/commit— stage, message, verify, commit/verify— 6-phase verification pipeline/feedback— capture friction as memory or CLAUDE.md update/debug— systematic debugging workflow/feature— end-to-end feature implementation/refactor— safe refactoring with verification at each step
Hooks: 4 total.
- PostToolUse → ESLint after every edit
- PostToolUse → TypeScript check after
.ts/.tsxedits - PreToolUse → Block writes to
.envfiles - Stop → Save session summary for continuity
MCPs: 3 total.
- Supabase — database operations and migrations
- GitHub — PR creation and issue management
- Context7 — documentation lookup for libraries
Memory: ~15 items. My role, my preferences, project context, feedback corrections from past sessions.
Your blueprint document
The final deliverable for this lesson is not a config file. It is a reference document for yourself.
Claude reads your actual files — CLAUDE.md, skills, settings.json. It does not need a summary. But you do. A blueprint.md that captures what you have built and why helps you maintain the system, onboard teammates, and audit your own setup.
# My Agent Blueprint
## CLAUDE.md
- Location: CLAUDE.md (repo root)
- Lines: [count]
- Sections: Commands, Structure, Boundaries, Anti-patterns, Routing table
## Skills
- /deploy — Pre-deployment checklist
- /review — Code review with severity levels
- /verify — 6-phase verification pipeline
- [your other skills]
## Hooks
- PostToolUse: ESLint after edits
- PreToolUse: Block .env writes
- [your other hooks]
## MCPs
- [list your connected tools]
## Memory
- [summary of key memory items]
## Last audited: [date]
## Next review: [date + 1 month]This document is for you. Update it when you add or remove pieces. Review it monthly. It is the map of the system you have built.