This lesson is being recorded
Every video of this course is being uploaded to my YouTube channel. Subscribe so you do not miss it.
Bruno Ashford @brunoashfordSubscribeModule 14 · Lesson 14.1 · 14 min
Assembling the Full System
Integrate your CLAUDE.md, skills, subagents, connectors and routines into one coherent, documented system.
Use ← → to move. The marker highlights text and the pencil draws on the slide; navigation locks while you annotate. Undo removes the last mark, Esc exits.
1Inventory everything you built
Ask Claude to list every component across your project and personal folders, grouped by type. You will almost certainly find duplicates, old experiments and one skill you forgot existed.
Inventory my whole AI setup: the project CLAUDE.md, my personal ~/.claude/CLAUDE.md, every skill in .claude/skills and ~/.claude/skills, every subagent in .claude/agents, every custom command, hooks and permission rules in settings files, connected MCP servers, installed plugins and scheduled routines. Output one table per type with name, purpose, last modified date and whether it overlaps with anything else. Do not change any files.2Turn CLAUDE.md into a router
Your CLAUDE.md should not contain every detail. It should say who the company is, the universal rules, and where to go for each type of task: "for proposals, use the quote-builder skill", "for anything financial, delegate to the finance-analyst agent". Keep it short so it loads fast and stays readable. Details live in the skills.
3Draw your AI org chart
Map each subagent as a specialist role reporting to you, with the skills each one owns underneath. If one agent owns fifteen skills and another owns one, rebalance. If two agents do the same job, merge them.
From the inventory, draw my AI org chart as a simple indented text tree: me at the top, each subagent as a department, and the skills each one owns underneath. Then list any skill with no clear owner, any overlap between agents, and suggest the smallest change that fixes each issue.4Map what runs alone and what waits for a human
List every routine and every action that leaves the building (email, message, payment, publishing, deletion). For each, confirm there is a human approval step or a hook that blocks it without approval. This map is the safety spine of your department.
5Test the whole system end to end
Run one realistic week in a single day: the morning brief, a sales task, a report, a finance summary. Watch whether each task reaches the right skill or agent on its own. Every wrong turn is a routing line missing from CLAUDE.md or a skill description that is too vague.
By now you have built a lot of pieces. A company handbook, a handful of skills, some subagents, a connector or two, a morning brief, a weekly report. Here is the risk: pieces that work on their own but do not know about each other, like a company where every employee is brilliant and nobody has seen the org chart.
In this lesson we connect them. One router file that sends every task to the right place, one org chart that shows who does what, and one map of what runs automatically and what always waits for a human.
Key points
- Inventory first: you cannot organize what you cannot see.
- CLAUDE.md is a router, details live in skills.
- Every skill needs one owner agent, every agent a clear department.
- Every action that leaves the building needs an approval gate.
Avoid
- Stuffing every instruction into CLAUDE.md until it becomes unreadable.
- Keeping duplicate skills because deleting feels risky.
- Automating an outbound action without a gate because it worked in testing.
- Testing components one at a time and never the full flow.
An online homeware store had built 22 skills over three months. The inventory showed four versions of a product description skill and no owner for refunds. The owner merged the duplicates, created a customer care subagent that owned refunds and replies, and rewrote CLAUDE.md as a two-page router. Tasks that used to go to the wrong skill about one time in five now land correctly almost every time.