BOB Overview — What It Is & How It Works¶
What is BOB? How does the tech stack work? What problems does it solve? Start here for a complete conceptual foundation.
What is BOB?¶
BOB = Business Operating Brain
BOB is an AI-powered system that helps businesses run better by making intelligent decisions and automating work. It's like having a smart business advisor available 24/7 — but it's not a chatbot for small talk.
Key difference from ChatGPT¶
| ChatGPT | BOB |
|---|---|
| Generic AI chat tool | AI system designed specifically for business operations |
| Answers questions | Solves business problems: sales, finance, HR, strategy, operations |
| No memory of your business | Understands your company: customer lists, processes, decisions, history |
| Each conversation is separate | Learns from your vault: patterns, preferences, past work |
| "Talk to AI" interface | Integrated into your workflow: works inside tools you already use |
| One model (GPT) | Multiple AI models: Claude, OpenAI, Gemini — you choose |
What Problems Does BOB Solve?¶
BOB handles questions and tasks across the entire business:
| Business Area | Example BOB Tasks |
|---|---|
| Sales | Lead scoring, proposal generation, pipeline forecasting, objection handling |
| Marketing | Campaign planning, content creation, audience segmentation, competitive analysis |
| Operations | Process optimization, vendor evaluation, cost reduction, bottleneck identification |
| Finance | Forecasting, budget reviews, supplier negotiations, cash flow analysis |
| HR | Hiring decisions, compensation review, skill gap analysis, team structure |
| Strategy | Market analysis, competitive positioning, growth planning, risk assessment |
The pattern: Wherever you have business data + a decision to make, BOB can help.
How BOB Works (The Tech Stack)¶
BOB is built on four layers. Each layer does one thing:
┌─────────────────────────────────────────────────────────┐
│ LAYER 1: OBSIDIAN (Your Knowledge Hub) │
│ • All business data stored as markdown files │
│ • Notes about customers, deals, processes, decisions │
│ • Vault = your company's brain (organized, searchable) │
└─────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────┐
│ LAYER 2: ACP PLUGIN (Obsidian Integration) │
│ • "Agent Client Protocol" plugin in Obsidian │
│ • Lets you ask BOB questions without leaving Obsidian │
│ • Sees your vault data, reads/creates notes │
└─────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────┐
│ LAYER 3: CLAUDE CODE (AI Agent Engine) │
│ • Receives your question from ACP plugin │
│ • Searches vault for relevant data │
│ • Chooses the right "skill" to use │
│ • Runs that skill to generate your answer │
└─────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────┐
│ LAYER 4: ANTHROPIC API (The AI Model) │
│ • Claude, OpenAI, or Gemini model │
│ • Does the actual thinking and writing │
│ • Can switch providers without changing anything else │
└─────────────────────────────────────────────────────────┘
What Each Layer Does¶
Layer 1 — Obsidian (Knowledge Storage)
- Your vault is a folder full of markdown files
- Files are organized: /{company.clients}, /{company.root}, /{resource.root}
- BOB reads from this vault to understand your business context
- When BOB creates output, some workflows save it back to the vault during the run; review the output paths shown in the run to see where files landed
Layer 2 — ACP Plugin (The Interface) - ACP = Agent Client Protocol - It's a plugin that runs inside Obsidian - You type your question into ACP; it passes it to Claude Code - ACP also gives Claude Code access to your vault files
Layer 3 — Claude Code (The Agent) - Claude Code is an AI coding assistant from Anthropic - But BOB uses it as an agent — it can think, search, and take actions - When you ask BOB a question: 1. Claude Code reads your question 2. It searches your vault for relevant information 3. It picks a "skill" (a workflow) to handle your request 4. It runs that skill 5. It returns the answer to you
Layer 4 — The AI Model (The Brain) - The actual AI doing the thinking - Could be Claude (Anthropic), GPT-4 (OpenAI), or Gemini (Google) - You can switch without changing Layers 1–3 - Think of it like swapping out the engine in a car — the rest still works
The Data Flow: How a Question Becomes an Answer¶
Let's trace what happens when you ask BOB a question:
You (in Obsidian):
"Analyze our top 10 customers and recommend which ones to prioritize for upsell"
↓
ACP Plugin:
Captures your question, passes it to Claude Code
↓
Claude Code:
1. Searches your vault: "Find all customer files in /{company.clients}"
2. Reads recent deals, contact history, revenue data
3. Searches for a skill: "Is there an upsell-strategy skill?"
4. Runs that skill with your customer data as input
↓
Claude Model (API):
Thinks through your customer data, scores them, ranks by upsell potential
↓
Claude Code:
Formats the result as markdown; depending on the workflow, either creates a new note in your vault directly or returns the draft for you to save
↓
ACP Plugin:
Shows the answer in Obsidian; if a path was written, it's listed in the run output
↓
You:
Review the analysis (and the output paths), edit if needed, ask BOB to save anything not yet saved
Why This Stack Matters¶
| Layer | Why It's There |
|---|---|
| Obsidian | All your data is yours, in files you own. No vendor lock-in. You can search, link, organize it your way. |
| ACP | You never leave Obsidian. No switching apps. BOB integrates into your existing workflow. |
| Claude Code | It's not just a language model — it's an agent that can think, search, and act. It can read your vault, pick the right skill, and execute it. |
| API | You can swap the AI model. If Claude is slow, switch to OpenAI. If you prefer Gemini, use that. Same system, different brain. |
What Are "Skills"?¶
A skill is a workflow that BOB can run. Think of it as a mini-program designed for a specific business task.
Examples:
- lead-qualifier — reads a lead's profile, scores them, decides if worth pursuing
- proposal-writer — interviews you about a deal, drafts a proposal
- market-sizing-analysis — researches a market, estimates opportunity size
- competitive-battlecard — compares your product to competitors
How Claude Code picks a skill: 1. You ask BOB a question 2. Claude Code reads your question 3. It thinks: "What skill would solve this?" 4. It loads that skill (a set of instructions + templates) 5. It runs that skill using your data 6. It returns the result
This vault includes a comprehensive library of skills, covering sales, marketing, operations, finance, HR, and more.
What You'll Install¶
A working BOB setup has:
✅ Obsidian — knowledge management app
✅ ACP Plugin — agent interface inside Obsidian
✅ Claude Code — AI agent engine (runs locally)
✅ This vault — a full library of business skills + your company data
✅ API keys — Claude/OpenAI/Gemini credentials (you choose)
When it's all installed: - You open Obsidian - You type a business question into the ACP panel - Claude Code searches the vault, picks a skill, runs it - You get an answer, review it, and the output paths the run wrote to (some workflows save during execution; if a path isn't shown, ask BOB to save the draft) - Done — no app-switching, no chat tabs, no email chains
Key Concepts to Understand¶
- Vault = Your Company's Brain
- Every decision, customer, deal, process is in markdown files
- BOB reads from this vault to understand your context
-
The more structured your vault, the better BOB's answers
-
Skills = Workflows
- Pre-built decision-making processes
- "Draft a proposal", "Score a lead", "Analyze a market"
-
Claude Code picks the right skill based on your question
-
Obsidian Never Leaves
- You ask BOB inside Obsidian via ACP plugin
- Answer appears inside Obsidian
-
No context-switching, no ChatGPT tabs, no email chains
-
Multiple AI Models
- Claude, OpenAI, Gemini all work
- Switch by changing one config file
-
Useful for cost, latency, or feature tradeoffs
-
Transparent & Auditable
- Some playbook runs create a TaskNote step log capturing the playbook, what was invoked, the original user input, and per-step output paths — enough to retrace what BOB did
- The
audit-trailskill compiles a readable record of what was run from the agent session logs — an activity trail, not a cryptographically tamper-proof or hash-chained ledger (it computes no hashes and does not chain entries). For formal tamper-evidence in client/Mode-3 work, layer on an out-of-band system (signed version history, an external WORM log) - All vault data stays in your markdown files — not a black box, you own the data
Next Steps¶
- Read through this guide carefully
- If any section is unclear, research it (links provided above)
- Verify your environment with the environment_checklist
- Refer back to this guide as you work through installation — the 4 layers will become concrete when you see them running