LegendaryPathwayOS

A model-agnostic business operating filesystem. 20 organizations. 20 departments each. One unified command structure.

F:\LegendaryPathwayOS\ · v2026

Section 01 / System Overview

Architecture

Plan

Obsidian

Thinking layer

→

Build

VS Code + Claude

Development layer

→

Process

Ollama

Local AI layer

→

Run

Docker

Deployment layer

Each organization follows an identical 20-department folder structure, ensuring consistency across all entities. The operating system is model-agnostic — it works with any AI model, CLI, or orchestration tool.

Organization Directory

#	Organization	Domain
01	Newmii	—
02	Rev Lawless Oliver Legend	lawlesslegend.com
03	Tuho Nipal	—
04	Enspireous	—
05	Lee Thomas IT / AGI R&D	leethomas.dev
06	Legendary Pathway	legendarypathway.com
07	Noble Trust Company	createyourtrust.org
08	Lex Liberorum Pax	lexliberorum.org
09	CHASE	—
10	WTF Is This WTF Does It Mean	—
11	M2M2 Messages To The Masses	messagestothemasses.org
12	Lexis Peptora	lexispeptora.org
13	CerebralFlex	—
14	MindBodyFlex	—
15	Maris Bakery	—
16	Lawless O Legend FREEDOM Clothing	—
17	Legends Market Place	—
18	Legend Hills	—
19	Watchback	watchback.org
20	Winters End Coffee Company	wintersendcoffee.com

Department Map (All 20)

#	Folder	Contents
00	Command	Org brain, priorities, LLM context, key links
01	Admin	Registration, licenses, insurance, vendors
02	Strategy	Vision, roadmaps, planning
03	Legal	Contracts, entity docs, compliance, IP
04	Finance	Budgets, banking, invoices, taxes
05	CRM	Contacts, leads, clients, pipeline
06	Sales	Offers, pricing, proposals, pitch decks
07	Marketing	Brand, campaigns, social, SEO, ads
08	Operations	SOPs, workflows, checklists, QA
09	People	Org chart, hiring, onboarding, training
10	Products_Services	Catalog, specs, pricing logic
11	Projects	Active / Planned / On_Hold / Completed / Templates
12	Website	Strategy, codebase, deployment
13	Content	Blog, newsletter, books, video, podcast
14	Tech_Data	Architecture, APIs, Docker, scripts
15	Partnerships	Strategic, affiliates, referrals
16	Meetings	Leadership, team, clients, partners
17	Documents	Reports, briefs, presentations, reference
18	Analytics	KPIs, dashboards, exports
19	Automations	Agents, prompts, workflows, scripts
99	Archive	Retired / completed items

Entry Protocol: When entering any org, read CLAUDE.md → README.md → INDEX.md → 00_Command/ → then work.
When entering a project: CLAUDE.md → README.md → project.yaml → then work.

Section 02 / Infrastructure

Infrastructure

Hetzner VPS (Primary)

Ubuntu 24.04 — Docker, Traefik reverse proxy, CRM, client portals, LP Secure VoIP.

SSH: root@5.161.206.175
Docker Compose orchestration
Traefik handles SSL + routing
CRM, portals, LP Secure all live
WireGuard VPN endpoint

STAYS — Production Core

Hetzner GEX44 (GPU)

ADDITIONAL server — ~$184/mo dedicated GPU for AI workloads.

Long-form video generation
Chatterbox TTS (MIT, self-hosted)
Open-source AI pipeline
Unlimited output, no per-token cost

GPU — AI Pipeline

Hostinger

Static site hosting — Apache, SFTP deploy.

lawlesslegend.com
legendarypathway.com
SFTP deployment
Apache web server

Static Sites

Syncthing

3-device real-time sync — encrypted, peer-to-peer.

Windows workstation (F: drive)
Hetzner VPS
S26 Ultra mobile
Vault + project files in sync

Sync — 3 Devices

Total monthly cost: ~$210–220/mo for unlimited AI output, production hosting, GPU compute, and real-time sync across all devices.

Network Topology

// Device → Network → Server

Windows Workstation (F:\LegendaryPathwayOS\)
  │→ Syncthing → Hetzner VPS (Docker, Traefik, CRM)
  │→ Syncthing → S26 Ultra (mobile command)
  │→ WireGuard → Hetzner VPS

Hetzner VPS
  │→ Traefik → *.legendarypathway.com
  │→ Traefik → secure.legendarypathway.com
  │→ code-server (remote VS Code)

Hetzner GEX44
  │→ Chatterbox TTS
  │→ Open-source video pipeline
  │→ Local AI models

Hostinger
  │→ lawlesslegend.com (static)
  │→ legendarypathway.com (static)

Section 03 / Command Cheat Sheet

All Skills, Commands, MCPs & CLIs

◆ Design & UI 6 skills

▼

/impeccable

Production-grade frontend. Modes: craft (shape-then-build), teach (design context setup), extract (pull reusable tokens).

USE WHEN: Building any web page or component from scratch. Most comprehensive design skill.

/taste

High-agency frontend engineering. Tunable parameters: DESIGN_VARIANCE(8), MOTION_INTENSITY(6), VISUAL_DENSITY(4). Anti-slop enforcement.

USE WHEN: You want premium, non-generic UI with tunable design parameters.

/emil-design-eng

Emil Kowalski philosophy. Animation decisions, spring physics, Sonner principles, micro-interaction polish.

USE WHEN: Polishing animations and micro-interactions specifically.

/ui-ux-pro-max

67 styles, 96 palettes, 57 font pairings, 25 charts, 13 stacks. Full design system intelligence.

USE WHEN: You need style/palette/font selection or design system decisions.

/frontend-design

Anthropic official frontend design skill. Distinctive interfaces, avoids generic AI aesthetics.

USE WHEN: Building distinctive interfaces without the full /impeccable setup.

/polish

Pre-deploy quality gate. Reviews changed code for correctness, security, UX edge cases, completeness.

USE WHEN: About to ship code. ALWAYS run before deploy.

◆ Video & Image 10+ tools

▼

GEX44 GPU Server

Long-form video generation. Open-source pipeline, self-hosted on dedicated GPU. Unlimited output.

USE WHEN: Long-form video, batch processing, or any compute-heavy AI task.

Dreamina

Short clips FREE. Uses same Seedance 2.0 model as Higgsfield. Zero cost for short-form.

USE WHEN: Quick short clips without burning credits.

/nano-banana

Image generation via Gemini CLI. Blog images, thumbnails, icons, diagrams, illustrations, photos.

USE WHEN: Generating any static image or visual asset.

/make-video

End-to-end video pipeline. Concept intake → Claude prompt → Playwright auto-submit to Higgsfield → job URL.

USE WHEN: Full automated video creation flow.

/higgsfield

Master model selector. Kling 3.0, Sora 2, Veo 3.1, Seedance 2.0, Soul 2.0, DoP, and more.

USE WHEN: Choosing which video/image model to use.

/higgsfield-prompt

MCSLA formula for prompt construction. Structured prompt engineering for AI video generation.

USE WHEN: Writing or refining any Higgsfield prompt.

/higgsfield-cinema

Multi-shot storyboard, Director Panel, Cinema Studio 3.0, shot sequencing, color grading.

USE WHEN: Creating multi-shot videos, storyboards, or professional filmmaking workflows.

/seedance

Seedance 2.0 multimodal. Filmmaker-voice discipline that bypasses content filters on first try.

USE WHEN: Using Seedance 2.0 specifically.

+14 more Higgsfield sub-skills

higgsfield-audio, higgsfield-camera, higgsfield-image-shots, higgsfield-mixed-media, higgsfield-moodboard, higgsfield-motion, higgsfield-pipeline, higgsfield-recall, higgsfield-recipes, higgsfield-soul, higgsfield-style, higgsfield-troubleshoot, higgsfield-vibe-motion, higgsfield-apps, higgsfield-assist

USE WHEN: Specialized video tasks (audio, camera control, character consistency, troubleshooting, etc.)

◆ Audio & Voice 4 tools

▼

Fish Audio S2 Pro

Voice cloning. Beats ElevenLabs PVC in quality. Primary voice tool for all non-Winston tasks.

USE WHEN: Any voice cloning or high-quality voice synthesis.

Chatterbox TTS

Free bulk TTS, MIT license, self-hosted on GEX44. Unlimited text-to-speech at zero marginal cost.

USE WHEN: Bulk TTS, narration, or any volume voice work.

ElevenLabs

Winston phone voice ONLY. Do NOT use for anything else. Restricted to AI phone system personality.

RESTRICTED: Winston phone system only. Never use for other purposes.

Voxtral

REJECTED. CC BY-NC license, requires 28GB VRAM. Do not self-host.

DO NOT USE. Rejected due to licensing and hardware requirements.

◆ Content & Viral 8 skills

▼

/viral:discover

Multi-platform topic discovery. Find trending topics across platforms.

USE WHEN: Starting content research, finding what is trending.

/viral:angle

Contrast Formula angle development. Turn a topic into a unique, shareable perspective.

USE WHEN: Developing a distinctive content angle.

/viral:script

HookGenie hooks and scripts. Write attention-grabbing hooks and full scripts.

USE WHEN: Writing hooks, scripts, or social media copy.

/viral:analyze

Multi-platform analytics collection. Gather performance data across channels.

USE WHEN: Reviewing content performance metrics.

/viral:onboard

Agent Brain setup. Initialize the viral content agent with your brand context.

USE WHEN: First-time setup of the viral content system.

/viral:setup

Platform connections. Connect social media accounts and configure integrations.

USE WHEN: Connecting platforms to the viral system.

/viral:update-brain

Brain evolution. Update the agent brain with new data, performance insights, and strategy shifts.

USE WHEN: Evolving your content strategy based on results.

/goviralbro

Research the last 30 days of viral content. Broad trend scanning and inspiration gathering.

USE WHEN: General viral research across all platforms.

◆ Development Workflow 11 skills

▼

/commit

Smart commit messages. Analyzes changes and generates meaningful commit messages.

USE WHEN: Ready to commit code changes.

/commit-push-pr

Full commit + push + PR flow. One command from working code to pull request.

USE WHEN: Ready to ship code to the repository.

/code-review

Review a pull request. Structured code review with actionable feedback.

USE WHEN: Reviewing someone else's PR or requesting review of yours.

/feature-dev

Guided feature development. Codebase understanding and architecture-focused implementation.

USE WHEN: Starting a new feature with architecture consideration.

/writing-plans

Plan before coding. Create structured implementation plans from specs or requirements.

USE WHEN: You have a spec and need implementation steps before touching code.

/executing-plans

Execute with checkpoints. Follow a written plan with review gates at each step.

USE WHEN: You have a plan and need to execute it methodically.

/systematic-debugging

Scientific debugging. Hypothesis-driven bug investigation with persistent state.

USE WHEN: Encountering any bug, test failure, or unexpected behavior.

/test-driven-development

Tests first. Write tests before implementation code.

USE WHEN: Implementing any feature or bugfix. Tests come first.

/verification-before-completion

Verify before claiming done. Run verification commands and confirm output before success claims.

USE WHEN: About to claim work is complete, before committing or creating PRs.

/using-git-worktrees

Parallel branches. Create isolated git worktrees for feature isolation.

USE WHEN: Starting feature work that needs isolation from current workspace.

/dispatching-parallel-agents

Run tasks simultaneously. Dispatch 2+ independent tasks in a single session.

USE WHEN: Facing 2+ independent tasks with no shared state.

◆ Orchestration 9+ skills

▼

npx vibe-kanban

Visual Kanban for multiple AI agents. Each card = agent + branch + terminal. Worktree isolation.

USE WHEN: Running multiple Claude/Codex/Gemini agents on different branches simultaneously.

/gsd:progress

Check progress, show context, route to next action (execute or plan).

USE WHEN: Starting a GSD session or checking project status.

/gsd:plan-phase

Create detailed phase plan (PLAN.md) with verification loop.

USE WHEN: Planning the next phase of work.

/gsd:execute-phase

Execute all plans in a phase with wave-based parallelization.

USE WHEN: Ready to execute a planned phase.

/gsd:autonomous

Run all remaining phases automatically — discuss → plan → execute per phase.

USE WHEN: You want full autopilot execution of remaining work.

/gsd:new-project

Initialize with deep context. PROJECT.md creation and full context gathering.

USE WHEN: Starting a brand new project from scratch.

/gsd:debug

Persistent debug state across context resets. Systematic debugging within GSD framework.

USE WHEN: Debugging within a GSD project.

/seed:seed

Typed project incubator. Guided ideation through graduation into buildable projects.

USE WHEN: Incubating a new project idea from concept to plan.

+30 more GSD commands

discuss-phase, verify-work, new-milestone, complete-milestone, add-phase, insert-phase, remove-phase, health, stats, settings, cleanup, pause-work, resume-work, map-codebase, audit-milestone, and more.

USE WHEN: Deep project management within the GSD framework.

◆ Obsidian 4 skills

▼

/obsidian-cli

Read, create, search vault content. Manage notes, tasks, properties from CLI.

USE WHEN: Interacting with Obsidian vault from Claude Code.

/obsidian-markdown

Wikilinks, callouts, properties, embeds, frontmatter, tags. Obsidian-flavored markdown.

USE WHEN: Working with .md files in Obsidian syntax.

/obsidian-bases

Database views (.base files). Filters, formulas, summaries, table/card views.

USE WHEN: Creating database-like views of vault notes.

/json-canvas

Mind maps, flowcharts (.canvas files). Nodes, edges, groups, connections.

USE WHEN: Creating visual canvases, mind maps, or flowcharts.

◆ Utilities 11 skills

▼

/caveman

75% fewer tokens. Ultra-compressed communication while keeping full technical accuracy.

USE WHEN: You want to reduce token usage dramatically.

/autoresearch

Autonomous goal-directed iteration. Modify, verify, keep/discard, repeat on any measurable metric.

USE WHEN: Iterating on a task with a measurable success metric.

/dream

Memory consolidation. Scans session transcripts for corrections, decisions, preferences. Auto-triggers every 24hrs.

USE WHEN: Consolidating session learnings into persistent memory.

/brainstorming

Required before creative work. Explores intent, requirements, and design before implementation.

USE WHEN: Starting ANY creative work. Mandatory first step.

/defuddle

Clean markdown from web pages. Removes clutter and navigation to save tokens.

USE WHEN: Reading web pages — use instead of WebFetch for non-.md URLs.

/skill-creator

Create, modify, evaluate skills. Build new skills or benchmark existing ones.

USE WHEN: Building or evaluating Claude Code skills.

/hookify

Behavior hooks. Create hooks to prevent unwanted behaviors from conversation analysis.

USE WHEN: Setting up automated behavior guardrails.

/open-source

browser-use library reference. Agent, Browser, Tools config, sandbox deployment, LLM models.

USE WHEN: Writing Python code using browser-use library.

/revise-claude-md

Update CLAUDE.md with learnings from the current session.

USE WHEN: Session produced insights that should persist.

/simplify

Code quality review. Check for reuse opportunities and efficiency improvements.

USE WHEN: Reviewing code for quality and maintainability.

/security-review

Security audit. Review pending changes for vulnerabilities and security issues.

USE WHEN: Before merging code that touches auth, data, or external APIs.

◆ MCP Servers 3 servers

▼

Figma MCP

Read designs, take screenshots, manage Code Connect mappings, search design systems, create FigJam diagrams.

USE WHEN: Working with Figma files, design-to-code workflows, or FigJam boards.

Gmail MCP

Search, read, draft emails. Gmail API integration for email management.

USE WHEN: Managing email from within Claude Code.

Google Calendar MCP

Calendar management. Create, read, update calendar events.

USE WHEN: Scheduling or checking calendar from within Claude Code.

◆ CLI Tools 6 tools

▼

npx vibe-kanban

Visual Kanban board for managing multiple AI agents, branches, and terminals in parallel.

USE WHEN: Multi-agent orchestration with visual tracking.

npx skills search|add|ls

Vercel skills registry. Search community skills, add new ones, list installed skills.

USE WHEN: Finding and installing new Claude Code skills.

claude

Claude Code CLI. Interactive AI-powered development assistant.

USE WHEN: Running Claude Code from terminal.

ollama

Local AI models. Run open-source LLMs locally for health checks and small tasks.

USE WHEN: Offloading small tasks from Claude API to save tokens.

GitHub CLI. Issues, PRs, checks, releases, and repository management.

USE WHEN: Any GitHub operation from the terminal.

docker / docker compose

Container management. Per-project Docker configurations, not per-org.

USE WHEN: Building, deploying, or managing containerized services.

Section 04 / Workflow

Building a New Website

Complete step-by-step flow from concept to production deploy.

/gsd:new-project

Deep context gathering. Creates PROJECT.md with full requirements, constraints, and goals.

/brainstorming

Explore intent and requirements. Mandatory before any creative work. Uncovers hidden assumptions.

/writing-plans

Implementation plan. Structured steps, dependencies, and milestones from the gathered context.

/impeccable teach

Set up design context. Establish the design system, tokens, and component patterns for the project.

/taste or /impeccable craft

Build with premium design. Shape-then-build approach with anti-slop enforcement and tunable parameters.

/emil-design-eng

Polish animations. Spring physics, micro-interactions, Sonner principles. The invisible details.

/polish

Pre-deploy quality gate. Correctness, security, UX edge cases, completeness review.

/verification-before-completion

Verify everything works. Run verification commands and confirm output before claiming done.

/commit-push-pr

Ship it. Commit, push, and open a PR in one flow.

npm run build && npm run deploy:production

Build and deploy to production. All sites deploy to live — never localhost.

Rule: All deployments go to live site. Never run localhost. Hostinger for static sites, Hetzner VPS for Docker services.

Section 05 / Workflow

Planning & Strategy

Structured planning flow using GSD orchestration.

/gsd:new-project or /gsd:new-milestone

Initialize project or milestone. Deep context gathering and PROJECT.md creation.

/gsd:discuss-phase

Gather context through adaptive questioning. Use --auto to skip interactive questions.

/gsd:plan-phase

Create detailed PLAN.md with verification loop. Structured phases, tasks, and acceptance criteria.

/gsd:execute-phase

Execute with wave-based parallelization. Multiple agents working on independent tasks simultaneously.

/gsd:verify-work

Validate built features through conversational UAT. Ensure goals are met before closing.

Workflow Hierarchy: Plan Mode is the default. Superpowers for parallel work. GSD is fully retired as of 2026-04-16 — the /gsd: commands remain but are used within the broader Plan Mode framework.

Section 06 / Workflow

Marketing & Content

Full content pipeline from discovery to performance tracking.

/viral:onboard

Set up the Agent Brain with your brand context, voice, and target audience.

/viral:discover

Find trending topics across all platforms. Multi-platform scan for what is resonating now.

/viral:angle

Develop content angles using the Contrast Formula. Turn trending topics into unique perspectives.

/viral:script

Write hooks and scripts with HookGenie. Attention-grabbing openings and full content scripts.

/nano-banana

Generate visual assets. Thumbnails, blog images, social graphics, illustrations via Gemini CLI.

/make-video or Dreamina

Create video content. /make-video for premium Higgsfield pipeline, Dreamina for free short clips.

/viral:analyze

Track performance across platforms. Gather analytics data for strategy refinement.

/viral:update-brain

Evolve strategy. Feed performance data back into the Agent Brain for continuous improvement.

Section 07 / Workflow

Video Production Pipeline

Three-tier pipeline: GEX44 (long-form) + Dreamina (short clips) + Higgsfield (paid premium).

/higgsfield-prompt

Concept → MCSLA formula. Structure the prompt with Motion, Camera, Subject, Lighting, Atmosphere.

/nano-banana

Generate image references. Create still frames, concept art, or reference images for video generation.

Dreamina (FREE)

Short clips using Seedance 2.0. Same model quality as Higgsfield at zero cost for short-form content.

GEX44 open-source pipeline

Long-form video on dedicated GPU. Self-hosted, unlimited output, no per-generation cost.

/make-video

Premium generation via Higgsfield. Full automated pipeline: prompt → Playwright → submit → job URL.

Fish Audio S2 Pro or Chatterbox TTS

Voice layer. Fish Audio for cloning quality, Chatterbox for bulk TTS (free, MIT, GEX44-hosted).

/higgsfield-cinema

Assembly. Multi-shot storyboard, Director Panel, shot sequencing, color grading, final cut.

Tier 1: Free / Self-Hosted

Dreamina — Short clips (Seedance 2.0)
GEX44 — Long-form (GPU pipeline)
Chatterbox — Bulk TTS (MIT)
/nano-banana — Image gen (Gemini)

$0 per generation

Tier 2: Premium / Paid

Higgsfield — Kling, Sora 2, Veo 3.1
Fish Audio S2 Pro — Voice cloning
/make-video — Automated pipeline

Credit-based

Restricted / Rejected

ElevenLabs — Winston ONLY
Voxtral — Rejected (CC BY-NC, 28GB)

Do Not Misuse

Section 08 / Workflow

Private Off-Grid Network

Mobile command center architecture for sovereign operations.

S26 Ultra

Mobile Command Center

Thin client to the full stack. Syncthing for vault sync, Termux for terminal, code-server for remote development.

Mobile Device

Cape Carrier

IMSI Rotation

AT&T backbone with IMSI rotation. $99/mo. Privacy-first mobile carrier.

$99/mo

Mullvad VPN

No-Identity VPN

Monero payment, no email required. True no-log, no-identity VPN service.

Monero Payment

WireGuard

VPS Tunnel

Direct encrypted tunnel to Hetzner VPS. Fast, modern VPN protocol.

To Hetzner VPS

Syncthing

Vault Sync

Encrypted peer-to-peer sync across all 3 devices. No cloud intermediary.

3-Device Sync

code-server

Remote VS Code

v4.115.0 on Hetzner. Full VS Code in the browser with Claude Code CLI access.

Browser IDE

Critical: Termux must be installed from F-Droid, NOT Google Play. The Play Store version is abandoned and outdated.

Network Stack

S26 Ultra
  │→ Cape Carrier (IMSI rotation, AT&T backbone)
  │→ Mullvad VPN (Monero, no email)
  │→ WireGuard tunnel → Hetzner VPS
  │→ Syncthing ←→ Windows Workstation
  │→ code-server (browser VS Code + Claude CLI)
  │→ Termux (F-Droid) for local terminal

Section 09 / Tool Selection Guide

When to Use What

Interactive decision cards. Find the right tool for the job.

Planning

Ultrathink (Extended Thinking)

USE WHEN: Complex architecture decisions, multi-step reasoning, system design choices.

/writing-plans

USE WHEN: You have a spec and need structured implementation steps before touching code.

/gsd:plan-phase

USE WHEN: Working within GSD project structure. Creates PLAN.md with verification loop.

Multi-Agent

npx vibe-kanban

USE WHEN: Running multiple Claude/Codex/Gemini agents on different branches simultaneously. Visual board + worktree isolation.

/dispatching-parallel-agents

USE WHEN: 2+ independent tasks in a single Claude session. No shared state needed.

/subagent-driven-development

USE WHEN: Executing plans with parallel independent tasks within current session.

Design Quality

/impeccable

USE WHEN: Building any frontend from scratch. Has craft/teach/extract modes. Most comprehensive design skill.

/taste

USE WHEN: You want tunable design parameters (variance, motion, density). Anti-slop enforcement.

/emil-design-eng

USE WHEN: Polishing animations specifically. Spring physics, Sonner principles.

/frontend-design

USE WHEN: Quick frontend work without the full /impeccable setup. Anthropic official.

/ui-ux-pro-max

USE WHEN: Need style/palette/font selection. 67 styles, 96 palettes, 57 font pairings.

/polish

USE WHEN: Code is done, about to deploy. Pre-deploy quality gate. ALWAYS run before shipping.

Content & Marketing

/viral:* suite

USE WHEN: Creating social media content, discovering trends, writing hooks/scripts. Full content pipeline.

/brainstorming

USE WHEN: Starting any creative work. REQUIRED first step. Explores intent before implementation.

/autoresearch

USE WHEN: Iterating on any task with a measurable metric. Autonomous goal-directed improvement.

Skills Management

/using-superpowers

USE WHEN: Starting any conversation to discover available skills.

/skill-creator

USE WHEN: Building or evaluating skills. Create, modify, benchmark.

npx skills search|add

USE WHEN: Finding and installing community skills from the Vercel registry.

Code Quality

/polish

USE WHEN: Pre-deploy quality gate. Correctness, security, UX edge cases.

/simplify

USE WHEN: Code reuse and efficiency review. Find opportunities to reduce complexity.

/security-review

USE WHEN: Security audit. Before merging code that touches auth, data, or APIs.

/systematic-debugging

USE WHEN: Bug investigation. Hypothesis-driven, scientific approach with persistent state.

/verification-before-completion

USE WHEN: Before claiming done. Evidence before assertions, always.

Token Efficiency Rule: Use /caveman for 75% fewer tokens. Offload health checks and small tasks to local Ollama models (Gemma, Llama) instead of burning Claude API tokens.