Stop managing your AI. Start building.
You open Claude Code. You explain the project. Again. Claude starts coding before you finish the sentence. You tap yes on a diff you half-read. Three sessions later, the codebase has patterns you don't recognize.
You switch to Codex for a task. Start over. Explain everything. Again.
Your teammate opens the project. "What's the context here?" You spend 20 minutes on a call.
You come back after a week. Where were we?
You open Claude Code.
Type /plan.
Claude already knows what you're building, your stack, your constraints, where you left off.
It shows you the steps. Waits for you to say go.
You finish a feature. Type /ship.
Claude checks your branch status. Finds you're 2 commits behind.
Asks if you want to rebase. Handles it.
Runs a security check. Finds a console.log you left in.
Generates the commit message. Generates the PR description.
Asks before doing any of it.
Your teammate opens the project. Reads .project.md. 25 lines. Knows everything they need.
You come back after a week. Type /plan. Still knows.
You switch to Codex. Drop in the same .project.md. Still knows.
That's the difference.
CLAUDE.md ← behavior rules, loaded every session automatically
.project.md ← 25-line project brain, survives everything
skills/ ← slash commands inside Claude Code
├── /setup ← sets up .project.md, works on new and existing projects
├── /plan ← steps defined, you approve before code is written
├── /ship ← branch check → conflicts → security → commit → PR
├── /review ← code review against your project's constraints
├── /debug ← root cause before any fix is attempted
├── /ui ← interface + aesthetic direction, approved before code
└── /sync ← end-of-session brain update
hooks/
└── session-start.sh ← shows project state at every session open. Automatic.
You pay for every token. This is what that looks like:
| Tokens per session | |
|---|---|
| Most context systems | 1500–4000 (every session, before any work) |
| This package | ~400 (CLAUDE.md only, unless you run a skill) |
| With /plan or /ship | ~700 (context loads only when needed) |
Context on demand. Not context by default.
Every session, Claude operates under these rules:
- Ask before assuming. If a task has multiple interpretations, it lists them. Doesn't pick.
- Plan before building. Any task touching more than one file gets a step-by-step plan first. You approve it.
- Touch only what's needed. No cleaning up adjacent code. No adding unrequested features.
- You decide, it executes. No architectural decisions made alone. No files created outside task scope. No destructive commands without explicit approval.
- Short answers. No preambles. No summaries after the fact. No "I'll now...".
It's 25 lines. Plain text. Looks like this:
---
project: BillingDash
purpose: SaaS billing dashboard for indie developers
stack: Next.js, Stripe, Postgres, Vercel
constraints: idempotent webhooks, sub-200ms API responses
---
decisions:
2026-01-15: using Supabase Auth over custom — faster to ship, meets compliance
2026-02-03: polling over websockets for now — premature optimization risk
state:
current: webhook handler for failed payments
next: retry logic with exponential backoff
blocked:Any AI reads this cold and is oriented in one message. Any human reads this and understands the project without a call. No lock-in. No tooling. Just a file you own.
/plugin marketplace add champ18ion/claud18code
/plugin install claud18code@champ18ion/claud18code
Then open your project in Claude Code and type /setup.
New project — 5 questions, .project.md written, done.
Existing project — scans your codebase, pre-fills what it finds, you confirm. Under a minute.
Or install manually:
git clone https://github.com/champ18ion/claud18code
cp claud18code/CLAUDE.md claud18code/.project.md your-project/
cp -r claud18code/skills claud18code/hooks your-project/Day 1 → /setup five questions, context set, never repeat yourself
Every task → /plan steps defined, you approve, then Claude executes
Every commit/PR → /ship full git workflow, you make every decision
Any bug → /debug hypothesis first, root cause confirmed, then fix
Code review → /review critical issues flagged, nothing padded
Any UI work → /ui design read + memorable detail approved before code
End of session → /sync brain updated, project state preserved
.project.md is plain text. It works with Claude Code, Codex, Cursor, or whatever comes next.
You built the context once. You never rebuild it.
- Andrej Karpathy — whose observations about LLM coding pitfalls shaped how developers think about AI behavior. Our CLAUDE.md goes further with one rule his doesn't have: developer stays in charge.
- The token optimization community — caveman, antigravity, and others who proved output verbosity is waste. We built that assumption in from the start.
MIT License