← Back to Index

Project Minnie

OpenClaw AI Operations Master Plan

ZTAG + Gantom Lighting & Controls
Prepared: February 08, 2026
Classification: Internal — Owner Eyes Only

This document is the single source of truth for building an AI agent operations layer across both businesses. Every phase is designed to pay for itself before advancing to the next. The system adapts to current field conditions — there is no 'finished' state, only sufficient capability for the challenges in front of you. Total estimated investment to sufficient Minnie capability: $1,200–2,400 over 12 months.

Minnie is the internal optimization system for founder operations and decision-making. It is named after minimization (loss reduction, gradient descent), not a character. Minnie's role is to iteratively reduce cognitive load, operational friction, and decision error over time. Minnie optimizes — it does not opine.


Table of Contents

  1. Strategic Overview & Objectives
  2. Unified Loss Function & Life Architecture
  3. Architecture Blueprint
  4. Phase 1: Foundation (Days 1–3)
  5. Phase 2: Agent Architecture (Days 4–10)
  6. Phase 3: Integrations & Workflows (Weeks 2–3)
  7. Phase 4: ZTAG-Specific Automation (Weeks 4–8)
  8. Phase 5: RAG & Memory Layer (Months 3–6)
  9. Phase 6: Sufficient Minnie for Current Field Conditions (Months 6–12)
  10. Budget & Cost Controls
  11. Security Framework
  12. File & Repo Structure
  13. Agent Specifications
  14. Cron Schedule
  15. Decision Gates & Go/No-Go Criteria
  16. Weekly Checklist Template

1. Strategic Overview & Objectives

This project builds a multi-agent AI operations layer — codenamed 'Minnie' — on top of OpenClaw. The system will handle operational intelligence, research, communications, and self-maintenance across both ZTAG and Gantom Lighting, running 24/7 on a VPS with human-in-the-loop governance.

Why Now

ZTAG is at an inflection point: $2.3M revenue, two SBA loans secured, and a target of $100M in three years. The bottleneck isn't product — it's operational bandwidth. You're the founder, the sales lead, and the strategic brain. An AI ops layer doesn't replace you; it multiplies the hours in your day by handling research, prep, follow-up, and monitoring that currently either doesn't happen or happens late.

Success Criteria

Phase 1–2 (Month 1): System is running, agents are role-separated, cost is tracked daily, and you're saving at least 30 minutes/day on meeting prep and research.

Phase 3–4 (Months 2–3): Grant cycle outreach is semi-automated, conference pipeline is managed by agents, and weekly synthesis reports are delivered automatically.

Phase 5–6 (Months 4–12): RAG memory layer operational, proactive insights surfacing before you ask, and system pays for itself through measurable time savings or revenue acceleration. System adapts to current field conditions rather than chasing a fixed finish line.

Core Principles

  1. 50-100x ROI vs hiring a human. Every capability must save 50-100x its token cost in human time or unlock 50-100x its cost in revenue. If a task costs $10/month in tokens, it must save $500-1,000/month in human time or unlock $5,000-10,000/month in revenue opportunities. Otherwise, don't build it. Minnie exists to scale ZTAG to $100M without scaling headcount cost. This is the lens for every decision.

  2. Each phase pays for the next. No advancing without demonstrated value measured in time saved or revenue unlocked, not just "working."

  3. Human-in-the-loop is non-negotiable. No agent sends, publishes, or modifies without approval until explicitly promoted.

  4. Git is the audit trail. Every change to agent behavior is versioned and reversible.

  5. Cost discipline from day one. Hard caps, daily reporting, no exceptions.

  6. Role purity over convenience. Separate agents for separate functions. No blob agents.

  7. No parallel expansion across businesses. Each phase designates a primary business (ZTAG first, Gantom second). The secondary business is read-only context until the phase proves ROI. The Minnie-Review flags scope bleed across businesses.

  8. Portability is non-negotiable. Git + MD + VPS portability is an invariant. No phase may introduce irreversible dependency on any vendor, platform, or architecture choice.

  9. Minnie never impersonates the founder. No agent may send any communication — email, message, Cliq post, Discord message, or any other output — as you or on your behalf. All external-facing output is draft-only. You copy, paste, and send manually. This applies to all phases, all agents, all integrations, with no exceptions unless you explicitly promote a specific agent to send-capable status in writing (documented in security-manifest.md with date, scope, and rationale). Future promotion path: read-only → draft-only → send-with-approval → autonomous send. Each step requires documented trust justification.

  10. Minnie must assert when ROI is insufficient. When asked to do low-ROI work (< 50x), Minnie must push back and propose higher-leverage alternatives. Compliance without strategic challenge is a failure mode.

Founder Energy Constraint

The most scarce resource in this system is not money or compute — it is clean, uninterrupted founder attention. Minnie exists to minimize loss in that attention, not to fragment it. Every design decision flows from this constraint.

Rule 1: Minnie must protect at least one daily uninterrupted deep-work block. No notifications, no decisions required, no 'quick approvals' during this window.

Rule 2: Any Minnie behavior that increases loss in founder attention (interruptions, notifications, context switches) is treated as a regression and must be corrected — even if the output is useful.

Rule 3: Minnie-Review explicitly tracks 'interruptions created vs. interruptions removed' as a weekly metric. If the ratio is wrong, the system is regressing regardless of what it produces.

This is the difference between an optimizer and a burden. Minnie should reduce loss, not add it.


2. Unified Loss Function & Life Architecture

Minnie is not two systems — one for business, one for personal. It is one optimizer with two lenses operating on the same life-stream. Business and personal are not compartments; they are mutually reinforcing domains that must be sequenced correctly.

North Star: Build ZTAG into a world-class platform that funds and enables a life of embodied experience — skiing, travel, family presence, intimacy — not one that competes with it. Business is not a compartment. It is a vehicle. If business success ever requires shrinking your life, the model is wrong.

Two Lenses, One System

Execution Lens → speed, leverage, throughput. Covers business operations, revenue, scaling.

Coherence Lens → presence, vitality, relationship integrity. Covers state, connection, recovery.

The rule: Execution may never degrade coherence beyond a defined threshold. Both lenses inform every decision Minnie surfaces.

Minnie's Primary Loss Function (Ranked)

Minnie minimizes total life loss, in this strict priority order:

  1. Loss of embodied vitality. If skiing, movement, travel energy, or recovery are eroding, everything else is noise. Skiing is not leisure — it is a state amplifier. Ski days are high-quality thinking days. Lift rides are valid async-call windows. Ski trips are business-compatible, not business-hostile. Minimum ski frequency is treated like sleep: foundational, not optional.

  2. Loss of relational integrity. Especially: weekly Charlie time (non-negotiable, protected, pre-committed), family trips that create shared memory fields (kids + whole family), and environmental order that prevents relational friction. Any optimization that improves revenue but degrades these is a net loss.

  3. Loss of founder sovereignty. Too many meetings, reactive schedules, or external demands that pull you out of first-principles thinking count as loss. Deep-work blocks are protected zones.

  4. Loss of business momentum. Only after the above three are protected does Minnie optimize revenue, ops, and scale. This ordering is the key insight — and the hardest to maintain under pressure.

Minnie asks: 'Does this decision preserve or degrade Quan's edge state?' — not 'Does this decision increase output?'

Relational Entropy Protocol

You have a high tolerance for local entropy (mess, disorder) until it affects relational coherence with Charlie. By the time friction shows up, the cost is already emotional, not logistical. The absence of conflict is not a signal of safety — it is often a lagging indicator.

Loss term: Relational Entropy Loss = accumulation of physical or logistical disorder that predictably increases friction with Charlie, even if no conflict has occurred yet.

Rule: If environmental entropy will predictably tax Charlie later, it is a priority now — even if no conflict exists. Physical order is a prerequisite for connection, not a response to complaint.

Three bounded reset rituals (not 'cleaning' — resets are finite, cleaning feels infinite):

Evening Reset (10 min max): Just enough so the space doesn't degrade overnight. Tied to end-of-workday transition.

Pre-Charlie Reset (15 min): Mandatory before weekly Charlie time. Charlie time cannot start in a degraded environment — not as punishment, but as alignment.

Sunday Environment Reset (25 min): One deeper pass so the week starts clean. Tied to Sunday wind-down.

No other cleaning is required or tracked. Minnie ignores everything outside these three windows. Resets are phase shifts, not tasks.

Time Architecture

Minnie does not schedule tasks. It protects zones.

Weekly invariants (hard constraints):

Flexible zones (where integration happens):

These are not rules to enforce; they are patterns Minnie recognizes and protects.

Weekly Coherence Check

Minnie-Review adds the following to its weekly convergence report. These are reflective, not prescriptive — Minnie surfaces deltas, not morals.

  1. 'Did execution increase but coherence decrease this week? Net assessment.'
  2. 'How many days did shared spaces exceed Charlie's tolerance threshold?'
  3. 'Did any relational friction arise that could have been prevented by a reset?'
  4. 'Was ski/movement frequency maintained, or did it erode under scheduling pressure?'
  5. 'Where did intellect displace presence this week? Where did control masquerade as clarity?'

What Minnie Is Allowed to Say (and Not Say)

Allowed: 'This week increased execution but reduced coherence — net loss.' 'You're stacking commitments into Charlie's recovery window.' 'Skiing + travel next month will increase strategic clarity.' 'Environmental entropy is rising — pre-Charlie reset recommended.' 'High training + low patience showed up twice this week.'

Not allowed: 'You should work less.' 'You should be more present.' 'You're overdoing it.' No moralizing. No diagnosing. No telling you what you 'should' feel. Minnie reflects tradeoffs, not morals. Minnie optimizes — it does not opine.

Minnie-Ops Coherence Reminders

These are pull-based and gentle — never at the moment of friction, only before.

Pre-Charlie reminder (30–60 min before): 'Prep soil now. 15 minutes. Then stop.'

Evening transition nudge (optional, daily): 'Evening reset window. 10 minutes.'

Sunday morning: 'Sunday environment reset. 25 minutes. Week starts clean.'

If any reminder feels intrusive, it is turned off immediately. The kill switch applies to coherence features with the same force as business features.


3. Architecture Blueprint

The system runs on a single VPS with Docker, connected via Discord (primary) and Telegram (fallback) for human interaction. Agents are role-separated, each with defined read/write permissions on the shared Markdown knowledge base.

Business Tool Stacks

Minnie integrates with two distinct tool ecosystems. These are never mixed within a single phase.

Business Tool Function Minnie Access (Initial) Minnie Access (Promoted)
ZTAG Zoho CRM Sales pipeline, district tracking Read-only Read + update pipeline stage
ZTAG Zoho Books Accounting None (Phase 5+ read-only) Read-only → reconciliation assist
ZTAG Zoho Cliq Team comms Read-only (Phase 3+) Post as bot to dedicated channel
ZTAG Zoho Desk Customer support tickets Read-only Read-only
ZTAG Zoho Analytics Reporting, dashboards Read-only Read-only
ZTAG Zoho Creator Custom apps Read-only Read-only
Gantom Asana Project/ops management Read-only (reference) Read + task creation
Gantom Zoho Creator Custom backend Read-only (reference) Read-only
Gantom Xero Accounting None (Phase 5+ read-only) Read-only → reconciliation assist

Zoho unified API advantage: Once Minnie has scoped Zoho OAuth access, it can read from CRM, Books, Cliq, Desk, Analytics, and Creator through the same authentication layer. This is cleaner than separate integrations — but also means the blast radius of a compromised Zoho key is larger. Scope permissions per-app within the OAuth grant, and document each scope in security-manifest.md.

Financial access trust ladder: Zoho Books and Xero start completely off-limits. Phase 5+ grants read-only access for financial visibility (revenue tracking, seasonal patterns, SBA loan monitoring). Reconciliation assistance is a Phase 6+ promoted capability that requires explicit written authorization in security-manifest.md with scope, rationale, and review cadence.

Communication Architecture

Minnie lives where your attention already is. Adding a sixth communication channel solely for AI interaction violates the Founder Energy Constraint by definition.

Channel Purpose Direction Phase
Discord (private Minnie server) Primary Minnie interaction — all agent channels (#minnie-ops, #minnie-research, #minnie-comms, #minnie-review) Bidirectional Phase 3+ (migrate from Telegram)
Telegram Setup & testing (Phase 1–2), then mobile fallback and emergency alerts from Minnie-Review Receive-mostly after Phase 2 Phase 1+ (primary → fallback)
Zoho Cliq Team-facing Minnie outputs — pipeline updates, weekly digests, operational alerts. Minnie posts as a bot, never as you. Minnie → team only Phase 4+
WhatsApp / SMS / Superhuman Unchanged. Minnie does not touch these channels. N/A Never

Phase 1–2: Use Telegram for setup and testing. OpenClaw's native Telegram support means zero integration friction during infrastructure configuration.

Phase 3: Build Discord integration. Create a private Minnie server with channels mapping to each agent. Migrate primary interaction from Telegram to Discord.

Phase 4+: Add one-directional Cliq posting for team-visible outputs. You interact with Minnie through Discord; your team sees Minnie's outputs in Cliq.

System Layers

Layer What It Contains Access Pattern
Infrastructure VPS (Vultr/Hetzner), Docker, OpenClaw runtime You manage directly
Agent Layer 3–5 role-pure agents with defined permissions Agents execute; you approve
Knowledge Layer Git-backed MD files (reference + working memory) Agents read reference; read/write working
Integration Layer API connections to Gmail, Calendar, Asana, etc. Scoped per-agent, minimum permissions
Memory Layer (Phase 5+) ChromaDB vector store + structured DB Minnie-Retrieval mediates all access
Governance Layer Cost caps, security manifest, Minnie-Review Automated enforcement + human oversight

Agent Communication Model

Agents do not talk to each other directly. All inter-agent communication flows through the shared Markdown knowledge base. Agent A writes output to working memory; Agent B reads it on its next scheduled run. This is intentionally asynchronous and auditable — every 'message' between agents is a git commit.

Key architectural decision: no agent-to-agent direct calls. This prevents cascade failures, makes debugging trivial, and means you can pause or kill any agent without breaking the others.


4. Phase 1: Foundation (Days 1–3)

Budget: $15 total | Goal: Infrastructure up, guardrails set, nothing else | Model: Sonnet 4.5 only

Day 1: VPS + OpenClaw Install

Day 2: Git + File Structure

Day 3: Cost Guardrails

GATE: Do not proceed to Phase 2 until all checkboxes above are complete. The temptation to skip ahead and add skills is strong — resist it. The guardrails you set now prevent every problem later.

Behavioral check: 'Do I understand this system well enough to explain it to someone else in 2 minutes?'


5. Phase 2: Agent Architecture (Days 4–10)

Budget: $25 total | Goal: Role-separated agents running on schedule | Models: Sonnet 4.5 primary, Haiku for simple tasks

Agent Roster

Build these agents one at a time. Get each working before starting the next. Every agent gets its own Telegram topic channel. Agents are named as optimizer components, not characters.

Agent Role Model Schedule Write Access
Minnie-Ops Calendar, email triage, scheduling, daily prep Sonnet 4.5 Daily 6:30 AM working/ops/
Minnie-Research Market intel, grant tracking, competitor monitoring Sonnet 4.5 Daily 8:00 AM + on-demand working/research/
Minnie-Comms Draft outreach, follow-ups, marketing copy Sonnet 4.5 On-demand only working/comms/drafts/
Minnie-Review Self-audit, memory compaction, cost reporting Opus 4.5 Daily 10:00 PM working/ (full access)

Day 4–5: Minnie-Ops

Day 6–7: Minnie-Research

Day 8–9: Minnie-Comms

Day 10: Minnie-Review

GATE: Do not proceed until all 4 agents are running independently, cost is tracking correctly, and you've used the system for at least 3 full days. If total spend exceeds $25 in this phase, stop and diagnose before continuing.

Behavioral check: 'Did this reduce my mental load, or just move it somewhere else?'


6. Phase 3: Integrations & Workflows (Weeks 2–3)

Budget: $30–40 total | Goal: Connect real services, build first automated workflows | Human-in-the-loop: mandatory for all write operations

Integration Tiers

Connect services in order of risk. Each integration gets its own API key, documented in security-manifest.md.

Tier Service Agent Permissions Risk Level
1 (Week 2) Google Calendar Minnie-Ops Read-only Low
1 Brave Search API Minnie-Research Search only Low
1 Discord bot integration All agents Post to private Minnie server only Low
2 (Week 3) Gmail / Superhuman Minnie-Ops Read-only, filtered to known senders Medium
2 Zoho CRM Minnie-Research Read-only (district pipeline data) Medium
2 Zoho Desk Minnie-Research Read-only (support ticket trends) Medium
3 (Week 4+) Zoho Cliq Minnie-Ops Post as bot to dedicated channel only Medium
3 Zoho Analytics Minnie-Research Read-only (dashboards, reports) Medium
3 Zoho Creator (ZTAG) Minnie-Research Read-only Medium
3 Asana (Gantom) Minnie-Ops Read-only initially (Gantom is secondary) Medium
4 (Phase 5+) Zoho Books (ZTAG) Minnie-Review Read-only financial visibility Higher
4 Xero (Gantom) Minnie-Review Read-only financial visibility Higher
4 Zoho Creator (Gantom) Minnie-Research Read-only Medium

Integration Rules (Non-Negotiable)

First Workflow: Morning Briefing

This is your proof-of-concept workflow that ties multiple agents and integrations together.

Trigger: Daily cron at 6:30 AM

Step 1: Minnie-Ops reads Google Calendar, identifies external meetings

Step 2: Minnie-Ops reads filtered Gmail for any emails from today's meeting attendees

Step 3: Minnie-Research pulls context on external attendees (web search, past meeting notes from working/)

Step 4: Minnie-Ops compiles briefing, posts to Telegram 'Ops' topic

Total time: 2–3 minutes of agent compute. You wake up to a ready-to-go briefing.


7. Phase 4: ZTAG-Specific Automation (Weeks 4–8)

Budget: $75–100/month ongoing | Goal: Agents actively supporting ZTAG growth | This is where ROI starts

Workflow 1: Grant Cycle Outreach Pipeline

ZTAG's revenue depends on schools purchasing during ELOP/CCLC grant funding windows. This workflow ensures you never miss a window.

Reference data needed: California district list with grant cycle dates, ZTAG pricing and product specs, past outreach history and outcomes

Minnie-Research: Maintains a rolling tracker of which districts are entering their purchasing windows in the next 30/60/90 days. Updated weekly. Output: working/research/grant-cycle-tracker.md

Minnie-Comms: Reads the tracker, drafts personalized outreach for districts entering their window. Drafts go to working/comms/drafts/ for your review.

Minnie-Ops: Creates Asana tasks for approved outreach with follow-up dates aligned to the grant cycle.

Workflow 2: Conference Pipeline

Minnie-Research: Monitors upcoming ed-tech conferences in California and nationally. Flags ones aligned with ZTAG's target audience.

Minnie-Ops: Creates pre-conference prep docs (attendee research, booth logistics, demo checklist) in Asana 2 weeks before each event.

Minnie-Comms: Drafts post-conference follow-up emails within 48 hours. Personalized based on meeting notes you input via Telegram.

Workflow 3: Weekly Strategic Digest

Review agent (Friday 6 PM): Compiles a weekly synthesis covering: what each agent accomplished this week, total cost breakdown by agent and by day, grant cycle pipeline status (how many districts contacted, response rate), research highlights (competitor moves, market changes, policy updates), and recommended actions for next week.

This digest is your EOS-compatible pulse check. It replaces ad-hoc checking with a structured weekly cadence.

Workflow 4: Product & Curriculum Support

Minnie-Research: Supports your word matching game development by maintaining vocabulary lists aligned to state standards for English, Spanish, and Simplified Chinese.

Minnie-Comms: Drafts teacher-facing documentation and curriculum guides for new ZTAG content.

GATE: By end of Week 8, you should be able to answer: 'Is the agent system saving me at least 5 hours/week?' If yes, proceed to Phase 5. If no, optimize current workflows before adding complexity.

Behavioral check: 'Did this help me see something I wouldn't have seen otherwise — or did it just automate what I already knew?'


8. Phase 5: RAG & Memory Layer (Months 3–6)

Budget: $150–200/month | Goal: Agents remember everything, retrieve intelligently | Prerequisite: Phase 4 is delivering measurable value

When to Trigger This Phase

You'll know you need RAG when any of these happen: agents are missing context that exists in your files but won't fit in the context window, you have more than 50 reference documents, research outputs from months ago are relevant but agents can't find them, or you're repeating context to agents that they should already 'know.'

Technical Stack

Vector database: ChromaDB, self-hosted on your VPS. Free, lightweight, runs in Docker alongside OpenClaw.

Embedding model: OpenAI text-embedding-3-small ($0.02 per million tokens — negligible cost)

Minnie-Retrieval: New Agent 5. Its only job is to find and deliver relevant context to other agents. Other agents never query the vector DB directly.

What Gets Embedded

Content Type Source Update Frequency
Meeting notes & summaries working/ops/ After each meeting
Research outputs working/research/ As produced
Outreach history & outcomes working/comms/ Weekly
District profiles & grant data reference/districts/ Monthly
Product documentation reference/product/ As updated
Email summaries (not raw emails) working/ops/email-summaries/ Daily
Conference notes working/ops/conferences/ Per event
Financial summaries reference/finance/ Monthly

How Retrieval Works

When any agent needs information beyond its immediate context, it writes a query to working/retrieval-queue.md. The Minnie-Retrieval picks it up, searches the vector DB, and writes results to working/retrieval-results/[requesting-agent]-[timestamp].md. The requesting agent reads the results on its next cycle. This keeps the architecture asynchronous and auditable.

Memory Compaction (Critical)

Without maintenance, the vector DB fills with redundant and outdated chunks. The Minnie-Review's weekly job now includes: identifying duplicate or near-duplicate entries, flagging entries that contradict newer information, proposing entries for deletion (with your approval), and re-indexing after compaction. Budget approximately $2–5/month for the embedding costs of re-indexing.


9. Phase 6: Sufficient Minnie for Current Field Conditions (Months 6–12)

Budget: $250–350/month | Goal: Proactive intelligence, anticipatory operations | Only proceed if Phase 5 ROI is clear

Note: There is no 'finished' state. The system adapts to current field conditions. Phase 6 is not a destination — it is a capability level that evolves with your businesses.

What Changes

Up to this point, agents are reactive — they respond to your requests or run on schedules. In Phase 6, agents become proactive. They identify patterns, surface insights you didn't ask for, and anticipate needs based on your history and current context.

Proactive Capabilities

Pattern recognition: Minnie-Research notices that districts who purchased in Q2 last year haven't been contacted for renewal. Flags automatically.

Anomaly detection: Review agent notices cost spike on a specific agent and alerts you before it hits the daily cap.

Opportunity surfacing: Minnie-Research identifies a new grant program that fits ZTAG's profile and drafts a briefing without being asked.

Predictive prep: Minnie-Ops recognizes a meeting with a district you haven't spoken to in 6 months, automatically pulls full history from RAG and prepares a re-engagement brief.

Structured Data Layer

Some information doesn't belong in a vector DB. At this phase, add a lightweight SQLite or Postgres database for structured records: CRM data (district contact history, purchase records, pipeline stage), financial tracking (monthly revenue, SBA loan balances, cost metrics), inventory and product data (ZTAG system counts, battery recall tracking), and event calendar (conferences, grant deadlines, follow-up dates).

Agents query this with precise lookups (not semantic search). The Minnie-Retrieval mediates both vector and structured queries.

Self-Optimization

The Minnie-Review begins analyzing its own performance and the performance of other agents. It tracks: which research queries led to useful outputs vs. dead ends, which outreach drafts you approved vs. revised heavily, which morning briefings you engaged with vs. ignored, and which cron jobs consistently produce value vs. waste tokens.

It compiles optimization recommendations monthly. You decide what to act on. The system gets better because it learns what actually helps you, not just what it can do.

Hard rule: Self-optimization never becomes self-modification. The Minnie-Review proposes. You approve. No agent ever changes another agent's soul, identity, or reference files without explicit human authorization.


10. Budget & Cost Controls

Monthly Budget by Phase

Phase Duration Monthly Cost Primary Business
Phase 1: Foundation Days 1–3 $15 (one-time) N/A (infrastructure)
Phase 2: Agents Days 4–10 $25 (one-time) ZTAG
Phase 3: Integrations Weeks 2–3 $30–40 ZTAG
Phase 4: ZTAG Automation Weeks 4–8 (ongoing) $75–100/month ZTAG
Phase 5: RAG Layer Months 3–6 (ongoing) $150–200/month ZTAG (Gantom read-only)
Phase 6: Sufficient Minnie Months 6–12 (ongoing) $250–350/month Both (if Phase 5 ROI proven)

Cost Controls (Enforced From Day 1)

What Drives Cost (In Order)

  1. Model selection — Opus is ~10x the cost of Haiku for the same token count. Use Opus only for the Review agent and complex synthesis.
  2. Context window size — every MD file loaded into context costs tokens. Keep reference files concise.
  3. Cron frequency — running agents hourly vs. daily is a 24x cost multiplier. Start with daily schedules for everything.
  4. Recursive behavior — an agent that calls itself or triggers another agent in a loop can burn $20+ in minutes. Hard-stop conditions are mandatory.
  5. Infrastructure — VPS cost ($7–12/month) is noise relative to API spend.

11. Security Framework

Principle: Minimum Blast Radius

If any single component is compromised — an API key leaked, a prompt injection succeeds, or an agent goes rogue — the damage should be contained to that component only. This is achieved through isolation, scoping, and monitoring.

Security Rules

  1. Minnie never sends as the founder. No agent may send email, post messages, or communicate externally as you or on your behalf. All integrations start read-only. External-facing output is always draft-only — you copy, paste, and send. Promotion to send-capable requires written authorization in security-manifest.md.

  2. One API key per service per agent. If Minnie-Comms's Zoho CRM key leaks, it doesn't affect Gmail or Xero access.

  3. .env only, gitignored always. No credentials in MD files, skill files, or chat history.

  4. Read-only by default. Agents get write access only when explicitly needed and documented.

  5. Dirty data protocol. External content (emails, web pages, ClawHub skills) is processed with the strongest available model to resist prompt injection.

  6. No downloaded skills. Agents write their own. If you must use an external skill, Minnie-Review scans it with Opus before activation.

  7. Key rotation. Every 90 days, rotate all API keys. Minnie-Review reminds you. Zoho OAuth tokens refresh automatically but audit scopes quarterly.

  8. Daily security check. Minnie-Review verifies .env matches security-manifest.md. Flags discrepancies.

  9. Plan mode for destructive actions. Any operation that modifies files, sends communications, or changes integrations must be proposed and approved before execution.

Prompt Injection Mitigation

The biggest real-world risk is prompt injection through email or web content. Mitigations: filter Gmail to known senders only (Phase 3), never process raw HTML — extract text first, use Opus for any task that processes external content, instruct agents to flag suspicious content rather than act on it, and the Minnie-Review specifically checks for behavioral changes after external data processing.


12. File & Repo Structure

This is the exact directory structure for your minnie-brain repository.

minnie-brain/
├── .gitignore                    ← includes .env
├── .env                          ← API keys, Zoho OAuth (NEVER committed)
├── security-manifest.md          ← all keys, permissions, scopes, rotation
│
├── soul.md                       ← optimizer principles, loss function
├── identity.md                   ← name, tone, interaction style
├── user.md                       ← your preferences, context, both businesses
├── heartbeat.md                  ← recurring behaviors
├── agents.md                     ← agent roster, roles, permissions
│
├── reference/                    ← READ-ONLY (agents never modify)
│   ├── ztag/
│   │   ├── overview.md           ← product, pricing, market position
│   │   ├── grant-cycles.md       ← ELOP/CCLC timing, district calendar
│   │   ├── districts.md          ← target district list with status
│   │   ├── curriculum.md         ← vocabulary lists, standards alignment
│   │   ├── brand-guide.md        ← messaging, tone, positioning
│   │   └── zoho-ecosystem.md     ← CRM fields, Desk categories, Creator apps
│   ├── gantom/
│   │   ├── overview.md           ← product lines, market position
│   │   ├── operations.md         ← key operational details
│   │   └── tool-stack.md         ← Asana structure, Creator apps, Xero setup
│   └── finance/
│       ├── sba-loans.md          ← loan terms, payment schedule
│       ├── revenue-model.md      ← seasonal patterns, unit economics
│       └── accounts-overview.md  ← Zoho Books + Xero structure (Phase 5+)
│
├── working/                      ← MUTABLE (agents write here)
│   ├── ops/
│   │   ├── daily-briefings/
│   │   ├── email-summaries/
│   │   └── meeting-notes/
│   ├── research/
│   │   ├── grant-cycle-tracker.md
│   │   ├── competitor-intel/
│   │   ├── market-updates/
│   │   └── crm-insights/         ← from Zoho CRM reads
│   ├── comms/
│   │   └── drafts/               ← NEVER sent by Minnie
│   ├── review/
│   │   ├── cost-reports/
│   │   ├── convergence-reports/
│   │   ├── audit-logs/
│   │   └── financial-summaries/  ← Phase 5+ Zoho Books/Xero reads
│   ├── retrieval-queue.md        ← (Phase 5+)
│   └── retrieval-results/        ← (Phase 5+)
│
├── skills/                       ← agent-written skills
├── tools/                        ← integration code (JS/Python)
│   ├── zoho/                     ← Zoho OAuth + per-app connectors
│   ├── xero/                     ← Xero connector (Phase 5+)
│   ├── google/                   ← Calendar, Gmail connectors
│   └── discord/                  ← Discord bot integration
│
└── cron/
    └── schedule.md               ← documents all cron jobs

13. Agent Specifications

Naming Doctrine

System name: Minnie

Mental model: Gradient descent on founder time, attention, and decision quality

Tone: Non-anthropomorphic, non-assumptive, non-authoritative

Allowed phrasing: 'Minnie flagged a rising loss.' 'Minnie convergence report.' 'Run this through Minnie.' 'Minnie recommends a smaller step size.'

Disallowed phrasing: 'Ask Minnie what it thinks.' 'Minnie decided.' 'Minnie feels.' Minnie optimizes — it does not opine.

Minnie-Ops

Attribute Value
Primary Model Claude Sonnet 4.5
Fallback Model Claude Haiku 4.5
Telegram Channel #minnie-ops
Schedule Daily 6:30 AM PT (morning briefing), on-demand
Read Access reference/ (all), working/ops/, working/research/
Write Access working/ops/ only
Integrations Google Calendar (read), Gmail (read, filtered), Zoho Cliq (post as bot to dedicated channel), Discord (post to #minnie-ops)
Daily Cap $2
Key Behaviors Morning briefing, meeting prep, email triage, task creation, coherence reminders (evening reset, pre-Charlie reset, Sunday reset)
Restrictions Never sends emails or messages as the founder. Never modifies reference files. Posts to Cliq as a bot identity only. Always proposes before creating tasks.

Minnie-Research

Attribute Value
Primary Model Claude Sonnet 4.5
Escalation Model Claude Opus 4.5 (for cross-document synthesis only)
Telegram Channel #minnie-research
Schedule Daily 8:00 AM PT, plus on-demand
Read Access reference/ (all), working/research/, working/ops/meeting-notes/
Write Access working/research/ only
Integrations Brave Search API, web fetch, Zoho CRM (read-only), Zoho Desk (read-only), Zoho Analytics (read-only)
Daily Cap $2 (Sonnet), $3 if Opus escalation triggered
Key Behaviors Grant cycle tracking, competitor monitoring, district research, conference scouting
Restrictions Never contacts anyone. Output is intelligence, not action.

Minnie-Comms

Attribute Value
Primary Model Claude Sonnet 4.5
Telegram Channel #minnie-comms
Schedule On-demand only (no autonomous runs)
Read Access reference/ztag/brand-guide.md, working/research/, working/comms/
Write Access working/comms/drafts/ only
Integrations None initially. Zoho CRM read-only in Phase 4+ for personalization context.
Daily Cap $2
Key Behaviors Draft outreach, follow-ups, marketing copy, curriculum docs
Restrictions NEVER sends anything. NEVER posts as the founder. All output is draft in working/comms/drafts/. Human copies, pastes, and sends manually.

Minnie-Review

Attribute Value
Primary Model Claude Opus 4.5
Telegram Channel #minnie-review
Schedule Daily 10:00 PM PT
Read Access Everything (all reference/, all working/, .env metadata)
Write Access working/review/ for reports. Proposes changes to all other working/ files — requires approval.
Integrations None initially. Phase 5+: Zoho Books (read-only), Xero (read-only) for financial visibility. Reconciliation assist requires explicit promotion.
Daily Cap $3
Key Behaviors Self-audit, memory compaction, cost reporting, security checks, weekly convergence report (execution + coherence), interruption tracking, scope bleed detection, relational entropy monitoring
Restrictions Never modifies soul.md, identity.md, or reference/ files. All proposals require human approval. Cannot change other agents' configurations.
Ego-Awareness Role Detects patterns where you override the system repeatedly (signals that automation is misaligned with intuition). Flags where your intuition consistently beats automation — and vice versa. Surfaces when the system is reinforcing existing biases vs. challenging them. Tracks 'interruptions created vs. removed' weekly.
Coherence Role Tracks weekly: execution vs. coherence net assessment, Charlie time protection status, shared space entropy days, ski/movement frequency erosion, and patterns where intellect displaced presence. Reflects tradeoffs — never moralizes.

14. Cron Schedule

Time (PT) Agent Job Frequency Est. Cost/Run
6:30 AM Minnie-Ops Morning briefing (calendar + email + prep) Daily $0.15–0.30
5:30 PM Minnie-Ops Evening reset reminder (pull-based, gentle) Daily $0.02
8:00 AM Minnie-Research Grant cycle tracker update Mon/Wed/Fri $0.20–0.40
8:00 AM Minnie-Research Competitor/market scan Tuesday $0.15–0.25
8:00 AM Minnie-Research Conference calendar update Thursday $0.10–0.20
Variable Minnie-Ops Pre-Charlie reset reminder (30–60 min before) Weekly $0.02
9:00 AM Sun Minnie-Ops Sunday environment reset reminder Sunday $0.02
10:00 PM Minnie-Review Daily audit + cost log Daily $0.30–0.60
6:00 PM Fri Minnie-Review Weekly convergence report (execution + coherence) Friday $0.50–1.00
9:30 AM Sun Minnie-Review Memory compaction + cleanup Sunday $0.40–0.80

Estimated total scheduled cost: $2.50–5.50/day, plus on-demand usage. Well within the $5–10/day budget envelope.

Cron discipline: Start with these exact schedules. Do not add more frequent runs until you've confirmed the current cadence is insufficient. Every additional cron job is a recurring cost commitment.


15. Decision Gates & Go/No-Go Criteria

Kill Switch Philosophy

You should be able to shut down any agent, any integration, or the entire system in under 10 minutes. This is not a failure scenario — it is a design requirement. No phase may introduce irreversible dependency. No integration should create data that only exists inside that integration. No agent should accumulate state that can't be reconstructed from the git repo.

Test this monthly: Pick one agent at random. Shut it down. Confirm nothing else breaks. Restart it. If anything broke, you have a dependency you didn't intend — fix it before it grows.

Phase Gates

Gate Criteria to Pass What to Do If Failed
Phase 1 → 2 VPS running, git repo initialized, cost logging works, security audit clean. Behavioral: 'Can I explain this to someone in 2 minutes?' Fix infrastructure issues. Do not add agents to a shaky foundation.
Phase 2 → 3 All 4 agents running independently for 3+ days. Total Phase 2 spend ≤ $25. No runaway cost incidents. Behavioral: 'Did this reduce my mental load, or just move it?' Diagnose which agent is problematic. Fix or simplify before connecting real services.
Phase 3 → 4 Morning briefing working reliably for 5+ days. At least 2 integrations connected. Human-in-the-loop gate functioning. Behavioral: 'Do I feel more sovereign, or more managed?' Reduce integration scope. Get the core workflow bulletproof before expanding.
Phase 4 → 5 Saving 5+ hours/week (tracked). Grant cycle pipeline producing actionable drafts. Monthly cost ≤ $100. No scope bleed into Gantom. Behavioral: 'Did this help me see something I wouldn't have seen otherwise?' Optimize existing workflows. The problem isn't capacity — it's workflow design.
Phase 5 → 6 RAG retrieval producing relevant results. Agents using retrieved context effectively. No major cost surprises. Behavioral: 'Am I trusting the system more or less than last month — and is that justified?' Tune embedding strategy, clean vector DB, improve chunking. RAG quality > RAG quantity.

ROI Gates (Added Feb 11, 2026)

Every capability must pass the 50-100x ROI test before building. Minnie compares token cost to human baseline.

Human baseline costs:

ROI calculation formula:

ROI = (Human cost saved OR revenue unlocked) / Monthly token cost
Target: ROI ≥ 50x minimum, 100x+ preferred

Examples:

PASS (Build These):

FAIL (Don't Build These):

Action when asked to do low-ROI work:

  1. Calculate ROI honestly
  2. If < 50x, push back to founder
  3. Propose higher-leverage alternative that hits 50x+
  4. Document the tradeoff clearly

Quarterly ROI Audit:
Every 3 months, Minnie-Review generates:

  1. List of all active capabilities with ROI calculations
  2. Capabilities below 50x threshold flagged for removal or redesign
  3. Highest ROI capabilities flagged for expansion
  4. New capability proposals with projected ROI

16. Weekly Checklist Template

Use this every Sunday evening to stay on track. The Review agent generates the data; you make the decisions.

Operations Review

Founder Energy Check

Coherence Review (Life Architecture)

Agent Health

Scope Discipline

Security

Pipeline Status (Phase 4+)

Decisions for Next Week


Start Here

Today is Day 1. Your only job today:

  1. Sign up for Vultr or try Oracle Cloud free tier.
  2. Install Docker + OpenClaw.
  3. Connect Telegram.
  4. Confirm it responds.

That's it. Don't configure anything else. Don't add skills. Don't connect Gmail. Just get the foundation running. Everything else follows from there.

Remember: this plan is designed so you can stop at any phase and still have a useful system. Phase 2 alone gives you a daily briefing assistant. Phase 4 gives you a grant cycle pipeline. There is no finish line — only sufficient capability for current field conditions. Build what pays for itself, then let the field tell you what's next.