Getting good at Claude Code is one move: hand more of the doing to agents and keep more of your attention for judgment. We call the path the delegation ladder — six rungs from doing it yourself to running an autonomous fleet. Play with the ladder below, walk the two tracks, then steal the artifacts that get you up a rung.
A capable agent takes your task, plans it, and does the work. When the task is big, a lead agent breaks it into parts, hands each to a teammate agent working in its own isolated context, and merges the results. Some run in parallel. Some run while you sleep. You sit above the whole thing — specifying intent, setting constraints, reviewing output.
This is the design, not a metaphor. Claude Code is deliberately low-level and unopinionated: it hands you primitives and expects you to compose the workflow. Everything here is either a way to brief the team better, a way to add a specialist, or a way to let the team run without you. One question runs through all of it: am I doing work a well-briefed agent could do, or am I doing the briefing and the reviewing? The second is the job.
Every rung hands more of the doing to agents and frees more of your attention for judgment — but each one asks for more setup and more earned trust. Higher is not automatically better. Pick a rung to see the trade.
CLAUDE.md plus subagents scoped to narrow jobs. See the artifacts →Illustrative — the shape of the trade-off, not a benchmark. The rungs are real; the numbers just show which way each dial moves as you climb.
You need a GitHub account and a Claude subscription — no terminal, no local checkout. Every module works from a browser or the Claude mobile app. That is the point: you can build on the go.
claude --teleport
when you are back at a keyboard. The work lives in the cloud; you check in from anywhere.
vercel, netlify deploy,
flyctl deploy) or GitHub Actions — which matters higher up the ladder, when an
agent deploys for you. Only need it on your own machine? python -m http.server,
npx http-server, or your framework's npm run dev.
The rest of the value comes from turning one agent into a team you have configured deliberately — and letting it run without you. In order of leverage.
CLAUDE.md at your repo root is read at the start of every session: conventions,
architecture, preferred libraries, build and test commands, pre-commit checks. Keep it short and
concrete — a shorter file is followed more reliably and costs less context. Scope it at the
user, project, or local level. Claude also builds its own
memory across sessions.
Steal a starter CLAUDE.md →
/review-pr or
/deploy-staging, shared with your team through the repo. Loaded on demand, so they
keep standing context lean. Every recurring task you encode as a skill is judgment you only
capture once.
CLAUDE.md is guidance the model follows, a hook is a rule the system enforces.
Steal a format-on-edit hook →
/loop
repeats a prompt in a running session for quick polling.
Routines run on Anthropic-managed
infrastructure on a schedule or a trigger — even when your computer is off — created from the
web, desktop app, or /schedule. And headless mode
(claude -p) runs Claude
inside any script or pipeline. Steal a routine and a headless one-liner →
Enough theory. Paste these into a repository today — each is a concrete version of a lever above. They are deliberately small; the point is the shape, not the length. Copy, then adapt to your stack.
# Project: acme-web
## Stack
- Next.js (App Router) + TypeScript + Tailwind. Postgres via Prisma.
## Conventions
- Two-space indent. Named exports only, no default exports.
- Server components by default; add "use client" only when you need it.
- Prefer editing an existing component over creating a new one.
## Commands
- Install: npm ci
- Dev: npm run dev
- Test: npm test # run before every commit
- Lint + format: npm run lint && npm run format
## Guardrails
- Never hand-edit anything under db/migrations/.
- Never commit secrets. Config comes from .env (see .env.example).The context every session reads. This is the single highest return on the least effort in the whole tool.
---
name: reviewer
description: Reviews a diff for correctness and security. Read-only.
tools: Read, Grep, Glob
model: sonnet
---
You are a senior reviewer. Given a diff, surface only real defects:
correctness bugs, security holes, missing error handling. Skip style.
Return findings most-severe first, each with file:line and a one-line fix.
If you find nothing, say so — never invent issues.A specialist you delegate to, scoped to the minimum tools. Specialisation plus constraint is what makes a fleet reliable.
{
"hooks": {
"PostToolUse": [
{
"matcher": "Edit|Write",
"hooks": [
{ "type": "command", "command": "npm run format --silent" }
]
}
]
}
}Formats every file the moment it is edited — a rule the system enforces, not guidance the model may forget. The kind of guardrail a parallel fleet needs.
Every weekday at 08:00, for each open pull request:
- summarise what changed and flag anything risky,
- run the test suite and post failures as review comments,
- label the ones that are ready to merge.
Do not merge anything yourself.An agent that runs on a schedule while your computer is off. Create it from the web,
the desktop app, or /schedule.
tail -200 app.log | claude -p "Slack me if you see any anomalies"The whole philosophy in one line: Claude Code pipes, scripts, and chains like any other command-line tool.
Strip away the feature list and the differentiators Anthropic keeps returning to are consistent. The tool is intentionally thin so that you can be sophisticated.
In this order. Each one returns more than the effort it costs — and each moves you up a rung.
This page is the map. The newsletter article walks the whole route in prose, and the official documentation is always the current source of truth.
Read the full guide → Official docs →