← → to navigate · 1 / 14

Claude Code, set up like you mean it · 2026

The Agent
Harness.

An AI coding model on its own is a brilliant intern with amnesia. The harness — twelve parts of configuration, rules, and checks you wrap around it — is what turns it into a teammate you'd put on a real codebase. One slide per part. Framework concept inspired by Charlie Hills; the setup and words here are mine, from daily use.

George Moreas · georgemoreas.com/insights

Part 1 of 12

CLAUDE.md — standing orders.

  • A markdown file the model reads at the start of every session: your stack, your conventions, what 'done' means.
  • Two levels: ~/.claude/CLAUDE.md follows you everywhere; the repo-level file rules that project.
  • Every time you correct the model twice for the same thing, that correction belongs here — permanently.
~/.claude/CLAUDE.md

Part 2 of 12

The project brief — scope before code.

  • Run /init on a fresh repo and Claude drafts the project's brief from the codebase itself. Then you edit it — that edit is the real work.
  • Defines what's in scope, what's out, and what finished looks like, before a single file changes.
  • Commit it. A brief that isn't in the repo doesn't exist.
/init

Part 3 of 12

Memory — never repeat yourself.

  • Corrections and preferences persist across sessions instead of dying when the window closes.
  • Good memory hygiene turns one-off fixes into defaults: naming style, test framework, deploy quirks.
  • Review it occasionally — stale memory is worse than none.
/memory

Part 4 of 12

Permissions — autonomy is earned, not granted.

  • Default: the model asks before every action, and you review each diff. Start every new codebase here.
  • Scope trust per tool and per project in .claude/settings.json — which commands run free, which always ask.
  • Never blanket-approve on a repo with secrets, migrations, or deploy scripts within reach.
.claude/settings.json

Part 5 of 12

Plan Mode — align before it builds.

  • The model writes out its full plan as a document and waits. You annotate — 'don't touch the auth module' — before it edits anything.
  • Catches bad assumptions when they cost seconds, not hours.
  • Rule of thumb: anything touching more than three files goes through a plan first.
Shift+Tab → Plan

Part 6 of 12

Model routing — right brain for the job.

  • Not every task deserves the biggest model. Route cheap models to grunt work, save the strongest for judgment calls.
  • Start small, escalate on failure — the reverse wastes money silently.
  • Switch mid-session when the task changes shape.
/model

Part 7 of 12

Skills — workflows as commands.

  • Anything you've prompted three times becomes a saved skill you fire with one command.
  • A skill is a small markdown recipe: the steps, the standards, the output format — reusable forever.
  • My rule: 200 words of skill beats 2,000 words of re-explaining.
name/SKILL.md

Part 8 of 12

MCP — give the model hands.

  • Model Context Protocol connectors plug Claude into your real systems: database, browser, issue tracker — read and write, not copy-paste.
  • Start with two: a browser (so it can verify its own UI work) and your data layer.
  • Every connector you add deletes a category of 'let me paste that for you' from your day.
claude mcp add

Part 9 of 12

Hooks — the non-negotiables.

  • Deterministic code that fires on the model's actions. Not a suggestion — a gate.
  • A PreToolUse hook can hard-block edits to protected paths or force the linter before any commit.
  • CLAUDE.md is what you ask the model to respect. Hooks are what you guarantee.
PreToolUse

Part 10 of 12

Sub-agents — parallel, focused, clean.

  • Spin up workers that each get one job and a clean context — no drift from a bloated conversation.
  • Perfect for 'review these six files independently' work; they run in parallel and return one focused result each.
  • Defined once, reused everywhere in the repo.
.claude/agents/

Part 11 of 12

Agent teams — many hands, one task list.

  • Multiple agents coordinate off a single shared task list: owner, status, handoff.
  • The task list is the contract — it's what prevents three agents from solving the same bug three ways.
  • Still the frontier. Start here last, after the rest of the harness holds.
experimental

Part 12 of 12

The verify loop — check, score, then ship.

  • The model reviews its own work before you see it: run tests, run the linter, load the page, score against the definition of done.
  • Ask for it explicitly at first; then bake it into CLAUDE.md so it's automatic.
  • This one habit is the difference between reviewing work and re-doing work.
gather → act → verify → repeat

Assembly

The model is the engine. The harness is why it drives straight.

Assembly order that actually works: CLAUDE.md and permissions first — rules and boundaries. Then Plan Mode and the verify loop — alignment and quality. Then MCP, skills, hooks — leverage. Sub-agents and teams last — scale. Build it on a real project, not a toy: configuration only teaches you something when something is at stake.

Get in touch →Official docs: code.claude.com/docs →

Written by George Moreas — more insights · framework concept inspired by Charlie Hills.

Mental Models for the AI Economy

Get the next essay the day it publishes.

A few a month, written from real builds. Unsubscribe in one click.

George Moreas
About the author
George Moreas

Senior product manager and builder — 15+ years shipping enterprise products (AT&T, BMW Group), now running his own with AI. These essays are field notes from that loop: what AI actually changes about work, product, and the economics underneath. Based in South Florida.

Full background Work with me