03
Module 03 · Connecting Claude to the Vault

The Reasoning Engine

The vault is built. Now we give it a brain. Claude connects to your IP — and from this point forward, everything it produces is grounded exclusively in four years of your work.

Sprint Day
Day 04
Tools
Claude API · Anthropic
Est. Build Time
3 — 4 Hours
Scroll
01
Chapter 01

The Distinction That Changes Everything

There are two ways to use Claude. Most people use the first way — they open a chat interface, type a question, and get an answer drawn from everything Claude has ever been trained on. The output is competent. It's also generic, because it has no idea who you are or what four years of your creative thinking looks like.

This module sets up the second way. By connecting Claude to the vault through the API, we constrain it deliberately. Claude stops drawing on general knowledge and starts drawing exclusively on retrieved chunks from your IP. Every prompt it receives arrives pre-loaded with the most relevant sections of your methodology, your narrative, your voice. It reasons against that context. Nothing else.

The output stops sounding like the internet and starts sounding like Cyber Coastlines. That's not a stylistic preference — it's a strategic requirement. The E-Suite needs to think and write from inside the ecosystem, not from outside it.

Claude is not the intelligence. The vault is the intelligence. Claude is the reasoning layer that makes the vault speak. That distinction matters for everything we build from here forward.

Cyber Blueprint · Architecture Principle
02
Chapter 02

The Claude API

The Reasoning Layer
Claude API
AI Reasoning · E-Suite Intelligence · Vault Interface

The Claude API is how the ecosystem communicates with Claude programmatically — not through a chat interface, but through direct API calls that give full control over what Claude receives, how it reasons, and what it returns. This is not the same as using Claude.ai. The API is infrastructure.

The critical privacy distinction: Anthropic does not train on API data by default. Content processed through the API — your vault documents, your prompts, your methodology — does not feed back into Anthropic's training models. Your IP goes in, reasoning comes out, nothing is retained. This is one of the reasons the API was chosen over any consumer-facing AI tool for this ecosystem.

Cost is usage-based — there is no monthly subscription. Budget conservatively at $50–75/month for moderate E-Suite usage. The consultant should architect the retrieval layer efficiently so Claude only receives what it needs. Unnecessary token consumption is an avoidable cost.

03
Chapter 03

Meet the E-Suite

The E-Suite is the AI collective that runs this ecosystem. Five agents, each with a defined role, each powered by Claude reasoning exclusively against the vault. They are not chatbots. They are operational staff — built during this sprint, running indefinitely after it ends.

01
The Architect
Decision Intelligence
Eliminates ambiguity before it becomes a problem. When a strategic fork appears — two tools competing for the same seat, two approaches in tension — The Architect runs the decision against your stated principles and returns a recommendation. Not options. A position.
02
The Curator
Vault Intelligence & Maintenance
Monitors the vault on a 72-hour cycle. Picks up new content from Descript transcripts and Obsidian additions, runs ingestion through LlamaIndex, logs what changed, and delivers a weekly vault health report. The RAG never depends on a phone call to stay current.
03
The Navigator
Learning Experience Agent
Lives inside the Creative Companion course experience on Ghost. When a student asks a question or hits a checkpoint, The Navigator queries the vault and responds in your voice using your methodology. Not Claude's general knowledge — your frameworks, your language, your four years of thinking.
04
The Operator
Workflow Execution
Handles all repeatable execution. Form submitted — confirmation sent. Stripe payment cleared — Ghost access granted. Post scheduled — publishing triggered. Every workflow the consultant builds during the sprint gets handed to The Operator to run indefinitely.
05
The Analyst
Audience Intelligence
Every Monday morning, automated. Pulls from Fathom analytics, Ghost subscriber data, and Stripe transaction records. Returns one brief: what content performed, where the audience dropped off, what the revenue pattern looks like. No dashboard to check. Just the signal.
04
Chapter 04

The System Prompt Architecture

Every E-Suite agent runs on a system prompt — a set of instructions that defines its role, its constraints, its voice, and its relationship to the vault. These are not improvised. They are engineered documents that determine the quality of every output the agent produces for the life of the ecosystem.

Each system prompt has four components. First, a role definition — who this agent is and what it exists to do. Second, a vault instruction — an explicit directive that reasoning must be grounded in retrieved vault content only, never general knowledge. Third, a voice calibration — the tone, register, and specific vocabulary of the Cyber Coastlines ecosystem. Fourth, an output format — what the agent returns and how it structures its response.

All five system prompts get written, tested, refined, and committed to GitHub during this module. They are version-controlled like code because they are code — they are the logic layer of your AI staff.

"A well-engineered system prompt is indistinguishable from a well-trained employee. A poorly engineered one produces outputs that sound like they came from somewhere else entirely. Get these right before anything else runs on top of them."

Cyber Blueprint · Build Standard
05
Chapter 05

Build Checklist

[ ]
Create Anthropic API account and generate API key
Under the Cyber Coastlines LLC entity. Store the API key securely in GitHub secrets — never in plain text in any file.
[ ]
Configure API connection to vault retrieval pipeline
Every API call routes through LlamaIndex retrieval first. Claude never receives a prompt without vault context prepended.
[ ]
Write and test The Architect system prompt
Role, vault instruction, voice calibration, decision output format. Test with five real strategic questions before approving.
[ ]
Write and test The Curator system prompt
Vault monitoring logic, ingestion trigger conditions, health report format. Test with a simulated new content addition.
[ ]
Write and test The Navigator system prompt
Course context, student-facing voice, vault-only reasoning directive. Test with ten real student question types drawn from Creative Companion curriculum.
[ ]
Write and test The Operator system prompt
Workflow execution logic, trigger conditions, output confirmations. Test against the five primary workflow triggers.
[ ]
Write and test The Analyst system prompt
Data sources, report format, three-section brief structure. Run a test report with dummy data before connecting live sources.
[ ]
Run a full E-Suite integration test
All five agents queried in sequence. Outputs reviewed against the vault. Verify every response is grounded in ecosystem IP — not general Claude knowledge.
[ ]
Set up API usage monitoring and budget alert
Alert triggers at $50/month. This prevents unexpected cost spikes before the system is fully optimized.
[ ]
Commit all system prompts and API configuration to GitHub
Version-controlled. Every prompt change tracked. Module 03 is not done until GitHub is current.
Before Leaving This Module
Test Against Real Outputs

The E-Suite is not functional until every agent produces outputs that unmistakably sound like Cyber Coastlines. Run each agent against questions that have known answers in the vault — then check whether the output references the right frameworks, uses the right vocabulary, and reflects the right philosophy. If an agent sounds generic, the system prompt isn't calibrated yet. Fix it in this module, not after the system is live.

End of Module 03
The Engine Is Running.

Module 04 builds the destination — the public-facing home where your audience arrives, reads, and decides to stay.