My Claude setup

How I use AI.

Not as a toy. As a small team: skills for expertise, agents for independent digging, a vault for memory, and guardrails that keep it honest.

75skills, one per kind of problem

Playbooks I never have to name. The right one loads from what I say.

8agents for the heavy lifting

Separate Claudes, used only when the blank slate or the volume is the point.

1rulebook above everything

Truth, money, privacy and irreversible actions, binding on every session.

0made-up numbers

A gap stays a gap until I have a real figure.

The blueprint

One question in. The right expertise out.

This is the whole system on one page. I ask in plain words, a router picks the playbook, skills or agents do the work, the vault remembers, and the guardrails sit above everything.

Youask in plain words
Routerreads the topic, picks the playbook
75 skillsplaybooks loaded into this very chat
8 agentsblank-slate Claudes, only when earned
The vaultread before, write after
Answerplus a note of what was used
Guardrailsnever invent a numberask before anything irreversibleshow the pessimistic casekeep private things private

What it does for me

The habits that make it useful.

It knows me

A vault, not a blank page

Before helping, it reads the notes for that topic, so I never re-explain my business. When I mention something durable, it files it and updates old facts in place.

It picks the expert

I never name a tool

Ask about margins and a finance playbook loads. Ask about a launch and a go-to-market one does. It mentions which one it used so I learn the system.

It argues back

A red team on call

Before anything expensive, public or irreversible, a separate agent with no memory of my enthusiasm argues the case against it.

It stays honest

Gaps stay gaps

If a number is missing it says so. Nothing is invented, because the same notes feed my pitches and pricing.

It asks first

Drafts, not sends

Deleting, publishing, sending or spending always needs my go-ahead. The default output is a draft I press send on.

It shows its work

Assumptions on the table

Every money answer states its assumptions and shows the pessimistic case, not only the hopeful one.

How a question travels

Ask once. It finds its way.

01I ask in plain words

I never have to know a tool's name.Pricing, ads, legal, hiring, anything

02It routes by topic

Reads what I am really asking about.Silent, no menu

03A skill loads its playbook

Into the same chat, so it already knows everything I said.75 playbooks

04I get an answer and a note

A one-line mention of which skill was used.So I learn the system

05The vault learns

Anything durable gets filed for next time.Read before, write after

every conversation

Skills and agents

Two ways to bring in expertise.

The default

Skills

A skill loads a playbook into this conversation, where Claude already knows everything I have said. I never type a skill name. It is picked from the topic.

They cover:

Money and pricingGrowth and adsDistribution and salesOperationsLegal and complianceWriting and copyCreative directionStrategy and decisionsSpreadsheets, decks and PDFsLearning
Only when it earns it

Agents

An agent is a separate Claude with no memory of the chat. It is worth it only when the blank slate is a feature, or when the job is heavy independent digging.

Cheap and reversible: no committee. Expensive, public or irreversible: always a red-team first.

The eight agents

A team that argues, digs and builds.

Red teamArgues against a plan without being softened by our conversation. The blank slate is the value.
MentorAn outside read on me, uncontaminated by what I have just been arguing for.
Market researcherHeavy web research. Returns a conclusion, not a flood of sources.
AI scoutSweeps the tool landscape and finds the specific move it unlocks.
Bug hunterLong, systematic debugging loops. Runs alone, never in a committee.
Vibe coderBuilds things and reads lots of code that should not fill my chat.
Data analystChews through exports and files, and returns the decision.
First customersResearch-heavy digging on where the first buyers actually are.

The rules

The vault and the guardrails.

All my notes live in one organised vault, split by area of my work. The rules above it overrule every skill and every agent.

Rule 1

Read before, write after

Before helping, it reads the matching folder of my notes. When I mention something durable, it files it, updating old facts in place.

Rule 2

Never invent

A gap is marked as a gap, never filled with a plausible-sounding number. The vault feeds pitches and pricing, so a made-up figure is a real problem.

Rule 3

Ask before anything irreversible

Deleting, publishing, sending or spending needs my go-ahead first, and the default is a draft I press send on.

Rule 4

Brief the blank slate

Before an agent starts, I paste the actual facts into its prompt, not filenames. An unbriefed agent returns textbook advice.

Why it matters

This portfolio was built with it.

From the AI card recommender to the booking agent to the store itself, the same habit runs through everything: talk to real people, build a rough version fast, and keep a human in the loop until it earns trust.