LLM / Claude Code in a folder
- Prompt-driven
- Current repo or pasted context
- Session memory
- One-off answers or code edits
- User remembers what to ask next
Prompt → Model → Folder → Output
Private walkthrough · stakeholder education mode
From one-off LLM usage to an always-on company Brain.
This page explains the difference between using Claude Code or a normal LLM in a folder and operating with a source-aware Brain, Hermes runtime, harnessed tools, vectorized context, closed-loop reminders, and learning agents.
Brainsource truth first
01 · The contrast
The model is not the operating system.
Claude, Codex, and ChatGPT are powerful engines. The unlock is building the context, memory, harnesses, approvals, and feedback loop around them.
Brain-connected operating system
- Source-truth-driven
- Brain/vault read first
- Durable corrections and canon
- Harnessed tools + verified execution
- System detects loops and compounds
Source → Brain → Hermes → Harness → Action → Write-back
02 · Brain-first principle
Information is read through the Brain before the answer.
The vault is not “notes.” It is the structured operating memory that makes every future agent less generic and more aligned.
Raw archiveTranscripts, exports, docs, screenshots, recordings.
InboxNew facts, feedback, corrections, worklogs, gaps.
CorrectionsCompounding behavior rules that every future run reads.
CanonCurated truth after human review.
SkillsProcedures for repeatable work.
What changes
Instead of pasting context into a model every time, the system checks standing truth, applies current rules, and writes new learning back to the Brain.
User askBrain lookupTool/model workVerified outputWrite-back
03 · Always-on infrastructure
The system has a permanent home.
A VPS keeps Hermes running continuously, so the operating layer can monitor, execute, schedule, verify, and deliver even when no laptop is open.
SurfacesTelegram · Slack · CLI · BrowserWhere humans ask, approve, and receive cards.
Always-on computeHetzner VPS · UbuntuStable server home for scripts, monitors, tools, git, and browser checks.
Agent runtimeHermesGateway, tool calling, skills, memory, profiles, scheduled jobs.
ContextBrain repo + archivesSource truth, corrections, project files, curated memory, evidence.
OutputsCards · Sites · Drafts · ReportsHuman-readable artifacts with approval and evidence boundaries.
04 · Hermes, harnesses, and models
What is a harness? It turns model intent into verified work.
Without a harness, an LLM describes work. With a harness, it can safely call tools, read files, use browsers, generate assets, poll jobs, and return evidence.
Hermes
The runtime layer that receives the request, loads skills/context, routes to models and tools, and enforces operating boundaries.
Harness
A controlled wrapper that validates parameters, runs the tool/API/browser/file action, captures results, and sends evidence back.
Models
Claude, Codex, GPT, Gemini, local models, or Nous-hosted tools become engines inside a larger operating layer.
Human intentHermes contextModel reasoningHarnessed toolEvidence-backed result
05 · Legible data
Vectorization helps, but it is not the Brain.
The durable pattern is raw source truth → structured records → source-aware chunks → hybrid retrieval → graph relationships → cited output → correction loop.
01Raw source truth
02Canonical records
03Chunks + metadata
04Vector + keyword search
05Graph relationships
06Rerank + evidence
07Cited action card
08Correction loop
Key teaching point
Vectors are indexes. The source of truth stays in the raw archive and curated Brain, so outputs can be checked, corrected, and improved.
06 · Operating intelligence method
The method turns messy company reality into aligned action.
This is the Brain version of the operating-methodology story: a simple way to understand how context becomes structure, structure becomes action, and action improves the system over time.
Messy work → clear roles → systems → accountable execution → feedback loop.
This is the same pattern behind practical systemization work: reality is captured, organized, assigned, measured, corrected, and improved.
Messy company context → structured Brain → harnessed agents → approved action → compounding intelligence.
The AI layer does not replace the methodology. It sits on top of it, reads from it, acts through it, and learns back into it.
Operating Intelligence Methodfrom context to trusted action
Educate through the pattern.
The viewer sees the flow from raw reality to source truth, goal alignment, agent work, human judgment, and compounding learning.
Use examples, not manuals.
Voice-agent calibration, client-risk surfacing, relationship follow-up, marketing intelligence, and website building make the system understandable through outcomes.
Keep trust visible.
Each loop shows source evidence, confidence, gaps, approval points, and what gets written back into the Brain.
Outside worldCapture layerBrain layerInterpretation layerAgent loopHuman gateCorrection back
06 · Goals and objective alignment
The operating system should keep the work tied to the company focus.
Beyond answering questions, the Brain layer should constantly relate activity back to goals, owners, dates, risks, and the next useful decision.
Company goalsFocus lensgrowth · profitability · client risk · hiring · marketing · AI ops
Team activityleader updates, calls, Slack, projects
Source evidenceGmail, Fireflies, HubSpot, Calendar, Brain
Recommendationfocus next, risk, blocker, opportunity
Human actionapprove, ask, delegate, suppress, monitor
“What is this person/function working on?”
Maps activity to goal relevance, progress signals, blockers, and suggested leadership questions.
“What needs to surface before it becomes a problem?”
Detects account/client risk from calls and source systems, then routes it with evidence and relationship context.
“Where is attention forming?”
Uses HubSpot, website, impressions, SEO/GEO, and content signals to recommend what marketing should improve next.
06 · Examples already built
Show the system through working loops.
Each example should teach the same closed loop: trigger → sources → Brain/context → agent work → human gate → output → write-back.
Voice Agent Calibration
Trigger: Yoni and Ofri test the live MultiplyMii voice agent and speak calibration feedback into the recording.
Flow: Recording + feedback → approved cost/model knowledge → prompt patch → website lab update → next test → Brain lock-in.
Why it matters: A customer-facing AI experience improves from real operating knowledge without letting public callers redefine company truth.
Human gate: Public callers cannot train it live. Durable changes require Hermes/Yoni review.
Open voice-agent lab08 · Recommendations and operating intelligence
The Brain should also say what to focus on next.
The experience should not only explain the architecture. It should show that the system can recommend priorities from what it already knows and keep work tied back to company goals.
Turn voice-agent calibration into a repeatable sales coaching loop.
The system already has recordings, prompt patches, pricing/model boundaries, and feedback packets. Next value is converting that into a reusable coaching/eval rail for sales and client-facing conversations.
Make every update answer “which objective does this move?”
Leader check-ins, client risk, recruiting activity, and marketing signals should all map back to company goals, owners, dates, and business value.
The correction loop is already real.
Feedback is captured, turned into rules, patched into live artifacts, and written back. That is the compounding behavior stakeholders need to understand.
A visible “closed-loop agent” demo.
Animate one reminder or relationship follow-up from source detection through approval and verified close. This teaches the operating model better than abstract architecture alone.
Use HubSpot + website attention to guide content and conversion work.
The system can help marketing move from impressions and page signals to grounded recommendations about what to improve next.
Do not let demos imply uncontrolled automation.
Every impressive example needs a human approval and evidence boundary. That is what makes the system trustworthy enough to scale.
09 · Agents as operating staff
Named loops are easier to train than generic bots.
Instead of one vague assistant, the operating system can have accountable agent roles with sources, outputs, and boundaries.
Captures new truth, corrections, availability, and worklog notes into Brain inbox with owner/date.
Prepares relationship follow-ups from email, calendar, calls, and Yoni voice patterns.
Detects unresolved loops and proposes close, keep-open, scheduling, or archive decisions.
Reviews voice-agent calls, identifies bad outputs, and turns feedback into prompt/knowledge/routing patches.
Turns call transcripts and source systems into compact team intelligence.
Keeps work tied to company objectives, owners, dates, progress, risk, and next moves.
Surfaces relationship/account risk from calls, email, CRM context, and leadership signals.
Reads HubSpot/website/SEO signals and proposes content, conversion, and visibility improvements.
Writes back what worked, what failed, and what should compound into future behavior.
10 · How to build the architecture
Teach the concept, then give the build path.
The final section should help a stakeholder move from understanding to implementation: what to set up, what to connect, where humans decide, and what to automate first.
11 · Glossary
Plain-English definitions for the technical layer.
LLM
The language model engine. Powerful, but by itself it only sees the current prompt, files, and session context.
Claude Code / Codex
Execution-focused coding agents. Excellent inside a repo/folder, but not the whole company operating system.
VPS
A rented always-on Linux machine. It gives Hermes a stable home when laptops are closed.
Hermes
The agent runtime that connects chat surfaces, models, tools, memory, skills, files, schedules, and approvals.
Harness
The controlled wrapper that turns model intent into validated tool calls with evidence, polling, files, and guardrails.
Brain / Vault
The durable source of truth: inbox, corrections, canon, project context, procedures, source-backed readouts, and operating rules.
Vectorization
A semantic index that helps retrieve similar meaning. Useful, but not the source of truth.
RAG / GraphRAG
Retrieval systems that find relevant source chunks and relationships so the agent can answer with context, citations, and gaps.