Philosophy: Claude as your operating team
What this is: the five ideas every other layer is built on. Why it matters: the files and scripts in this kit are just one implementation. If you understand the ideas, you can rebuild the system in any tool.
1. Treat Claude like a team, not a chatbot
A chatbot answers one question and forgets you. A team knows who you are, remembers what went wrong last month, follows house rules, has specialists, and keeps working while you sleep.
Operator OS gives Claude exactly those things:
| A good team has… | Operator OS layer |
|---|---|
| House rules everyone follows | Rules: CLAUDE.md files (doc 01) |
| Shared memory of past decisions and mistakes | Memory: one note per fact (doc 02) |
| Specialists and playbooks | Skills: reusable workflows (doc 03) |
| Access to the company's tools | Integrations: connectors, CLIs, browser (doc 04) |
| Night shift | Automation: scheduled jobs (doc 05) |
| A way of working together | Conventions: grill first, popups, parallel work (doc 06) |
| Separation between employers/clients | Multi-account + NDA boundaries (doc 07) |
2. Every incident becomes a rule
The system grows from mistakes, not from planning. When something goes wrong (Claude sends an email you wanted to review, a backup job dies silently, five agents collide in one folder), you don't just fix it. You write it down as a rule with the reason attached:
Rule: External email is draft-only. Why: an email went out with a wrong figure before the operator saw it. How to apply: create a draft, show it, the human clicks Send.
The why is what lets Claude judge edge cases instead of following rules blindly. After a few months, your rules file is a record of everything you've learned about working with AI.
3. Decide before doing
Most wasted AI work comes from starting before the goal is clear. So Claude grills you first: it maps the open decisions, asks them as multiple-choice popups in a few rounds, and only starts once you both agree. Facts it can look up itself, it looks up. Decisions are yours.
Two minutes of clicking answers saves an hour of redoing.
4. Load context lazily
Claude's context window is a budget. Loading everything about every project into every conversation makes it slower, more expensive and more confused.
So context is layered and loaded on demand:
- Always loaded: who you are, house rules, a one-line index of memories.
- Loaded when mentioned: a project's own
CLAUDE.mdand memory index ("when the user mentions Ridge Coffee, readridge-coffee/CLAUDE.mdfirst"). - Loaded when needed: a skill's full instructions, a reference doc.
5. Humans hold the irreversible buttons
Claude can draft, research, build, deploy to staging, and prepare everything. Some actions stay with you:
- Sending anything external (email, chat to clients, posts)
- Spending money (ads, purchases, budgets): Claude creates things paused
- Submitting (applications, forms, payments)
- Production systems you don't own
This is what makes it safe to give Claude broad permissions everywhere else. Speed comes from trust, and trust comes from clear limits.
The operator mindset
You stop doing tasks and start designing how tasks get done. Your job each week:
- Notice friction ("I explained this again", "that broke again").
- Turn it into a rule, a memory, a skill, or a job.
- Prune what's stale.
That loop is the whole framework. Everything else is detail.