Everything João can open during the demo — public links first, local Mac-mini dashboards clearly marked.
| What | Link | Status |
|---|---|---|
| Padawan Showcase | padawan-showcase.pages.dev | Live |
| Sim-game / Padawan HQ base | Section on this page | Live |
| OpenClaw Studio / Padawan HQ source view | http://localhost:3000 | Local Mac mini |
| ARTE.SANO Full Site Demo | arte-sano-demo.pages.dev | Live · client-ready POC |
| Woowind DESIGN.md / Agentic AI demo | woowind-demo.pages.dev/agentic-ai-demo | Live |
| Woowind Demo Store | woowind-demo.pages.dev | Live |
| Exploritori Site | exploritori.com | Live |
| Exploritori Spanish Blog | exploritori.com/es/blog | Live · 65/65 translated |
| Exploritori Admin | exploritori.com/admin | Live · 65 posts / 70 campaigns |
| Whale Tracker — Paulo | whale-tracker-nine.vercel.app | Live |
| Polymarket Bot Admin | http://127.0.0.1:5050/admin | Local Mac mini · bot controls |
The sim-game idea is no longer just a sketch. OpenClaw Studio already has a Padawan HQ command-centre view, and the underlying data model now has real modules, crons, deploys, social queues, and client demos behind it.
Local source: openclaw-studio/src/features/dashboard/. This is the practical base layer, already wired as a live dashboard panel.
Turn the command centre into a 90s management sim: Game Dev Tycoon + Pizza Connection + SimCity, but powered by real OpenClaw data.
Keep the current command-centre data model, then add a dedicated pixel/isometric layer that makes agents, modules, queues, and incidents visible at a glance.
projects/padawan-hq/references/Brand guidelines are no longer passive docs. They are executable context for agents, client demos, social assets, and production QA.
Each brand gets a DESIGN.md recipe: palette, typography, spacing, visual rules, avoid-list, and taste constraints.
Markdown stores intent; HTML stores layout truth. Agents reason from both instead of guessing visual structure from prose.
A live class demo shows Woowind’s DESIGN.md being used as executable context for an agent-generated campaign page.
ARTE.SANO now has a full design.md, asset library, 22-page replacement-site POC, mobile QA rules, logo hierarchy, and live Pages demo.
A local-first AI operating system running from a Mac Mini in Barcelona. João sets strategy. Monica coordinates follow-through. Brian executes, verifies, writes code, updates memory, and uses sub-agents/tools when they are actually useful.
OpenAI Codex through OpenClaw. Handles strategy, implementation, code review, deployment, memory, and operational judgement.
Hermes executive assistant layer. Turns João's intent into clear briefs, tracks work, requests evidence, and keeps Brian aligned with the current priority.
Daily competitor research + SEO analysis. Scrapes competitor blogs via Firecrawl, analyses search rankings.
Generates Instagram posts, Pinterest pins, and carousels using AI image generation (Gemini).
Posts to Instagram + Pinterest daily. Reads Sherlock's report to pick the most relevant content.
Polymarket is now treated as Monica's lane, not Brian's default operational noise. Brian only touches it when João explicitly asks.
Monitoring layer. Checks bot health, cron status, site uptime. Sends Telegram alerts.
Exploritori moved from raw content generation to a quality-controlled operating loop: research, refresh, translation, social distribution, live admin state, and deployment verification.
| Metric | Value | Notes |
|---|---|---|
| Published posts | 65 | Live pipeline count · 0 review drafts / 0 blockers |
| Spanish rollout | 65/65 | Final batch shipped 2026-07-01; stale translations repaired before publish |
| Social bank | 70 campaigns | 4 ready / 64 done / 2 held · no pending quality review |
| Google Trust Sprint | Active | New generic drafts paused; refresh proven articles first |
| Admin truth | Live JSON | Use exploritori.com data over dirty generated local files |
| Deploy target | Cloudflare Worker | exploritori Worker with assets, not a Pages project |
| Affiliate health | Clean | Amazon Associates link checks and OneLink kept under active QA |
The system is intentionally boring at the data layer: Markdown for memory, JSON for live public state, SQLite for structured stores, and explicit provenance before reporting private data.
Canonical workspace + vault. Recent cleanup moved intelligence into summaries and domain hubs instead of loading huge context by default.
vault/ — summaries, modules, knowledge hubsBRAIN/ — learnings, strategies, decisionsmemory/ — daily session logsMEMORY.md — long-term curated memorySQLite/JSON stores where they fit, with live APIs treated as truth when local generated files can drift.
exploritori.db — posts, social, analyticspipeline.json — live Exploritori statusfact_events — star schema P&LAll via CLI + API. No UI clicking.
The trading stack is still a useful proof of dashboards, reconciliation, and model-driven execution, but it is no longer Brian's normal lane. That boundary matters because the operating system now routes work by owner and context.
| Component | What it does | Data source |
|---|---|---|
| Polymarket Bot | Sports prediction model, dashboard, execution/reconciliation history | Monica-owned lane |
| Whale Tracker | SEC 13F superinvestor portfolio tracker by Paulo | Live Vercel app |
| Budget / Expense Ops | Subscription and operating cost tracking | Google Workspace + local scripts |
| Brian Rule | Do not monitor or repair Polymarket unless João asks | BRAIN prevention rule |
Brian's memory is files, but the retrieval discipline has matured: summaries first, small context windows, BRAIN lessons before work, and source provenance for private data.
Search summaries first, expand only when useful, keep retrieved notes lean, and avoid loading the whole vault by default.
Self-improvement: learnings.md, strategies.md, decisions.md, opportunities.md. Brian reads these before every task.
When João corrects Brian, the root cause and prevention rule are written down immediately so the mistake rate drops over time.
The first 15 days proved the machine could be built. The interesting part now is that it kept compounding into a real operating layer.
Local-first orchestration on one Mac Mini, with public surfaces deployed to Cloudflare/Vercel only where they need to be visible.
| Layer | Technology | Purpose |
|---|---|---|
| Hardware | Mac Mini (Apple Silicon) | All services run locally |
| AI Orchestration | OpenClaw | Agent management, cron, messaging, gateway |
| Primary Model | OpenAI Codex / GPT-5-class routing | Brian — strategy, orchestration, code, verification |
| Specialist Models | OpenAI, Gemini, local Ollama, tool-specific models | Reviews, bulk tasks, image/video/audio, embeddings |
| Image Generation | Gemini / image generation tools | Blog thumbnails, social images, client visual concepts |
| Web Framework | Astro + static HTML + Next/Vercel where needed | Exploritori, Woowind, ARTE.SANO, Whale Tracker |
| Hosting | Cloudflare Workers, Cloudflare Pages, Vercel | Project-specific deployment targets |
| Database | SQLite | All structured data |
| Knowledge | Obsidian (Markdown) | Vault, wikilinks, graph view |
| Embeddings | Ollama (nomic-embed-text) | Local vector search via LanceDB |
| Web Scraping | Firecrawl | Competitor research, content extraction |
| Social Posting | Buffer GraphQL API | Instagram + Pinterest scheduling |
| Analytics | Umami Cloud | Privacy-first web analytics |
| Trading | Polymarket CLOB API | Finance module proof, Monica-owned operational lane |
| Version Control | GitHub | 809+ Exploritori commits, plus supporting repos |
| Security | DefenseClaw (Cisco) | Skill scanning, CodeGuard, runtime protection |
| Self-Evolution | OpenSpace (HKU) | Self-evolving skills, cloud community (1.6K ★) |
| Search | webserp | 7 engines in parallel, no API keys required |
Brian does not just use tools; he turns repeated work into reusable skills, proposals, BRAIN rules, and cron workflows. The practical win is fewer repeated mistakes and less token waste.
| Skill Category | Examples | How We Use It |
|---|---|---|
| Web Scraping | firecrawl-search, firecrawl-scrape, firecrawl-crawl | Sherlock competitor research, content extraction |
| Google Workspace | gog (Gmail, Calendar, Drive, Sheets) | Email monitoring, Drive image uploads, spreadsheet tracking |
| Image Generation | Gemini / image tools / deterministic composition | Blog thumbnails, social media designs, Pinterest pins, client concepts |
| GitHub | gh CLI, gh-issues | Code deployment, issue tracking, PR management |
| Voice & Media | peekaboo, video-frames, openai-whisper | UI automation, transcription |
| Knowledge | memory search, vault summaries, oracle, skill-creator | Small-context retrieval, research, creating/updating reusable skills |
| Social | xurl (X/Twitter) | Reading tweets, monitoring trends |
| OpenSpace / Skills | delegate-task, skill-discovery, Skill Workshop | Use proven workflows, propose reusable skills, delegate only when useful |
| Communication | weather, apple-reminders, apple-notes | Context for daily operations |
Every task triggers a reflection: What worked? What was inefficient? What can improve? Results stored in BRAIN/learnings.md
The security layer is now governance as much as scanning: least-context retrieval, private-source provenance, safer cron ownership, and DefenseClaw in observe mode.
Skills and plugins are treated as supply-chain inputs: scan, inspect, propose, then apply only when appropriate.
Generated code gets security-aware review, especially for secrets, command execution, SQL, and public-facing endpoints.
Observe mode plus operating rules: exact account checks, no private fallback data, and no public/external actions without intent.
OpenSpace and Skill Workshop are available, but the useful pattern is selective evolution: promote repeated work into skills only when it reduces future friction.
OpenSpace provides external skill discovery and delegation when a task benefits from a specialist worker.
Corrections from João are captured immediately as prevention rules, then promoted into strategies or skill proposals when they repeat.
The main efficiency gain is not magic automation; it is not rereading giant contexts, not repeating broken recommendations, and not guessing when live data exists.
Every project in Padawan AI is a module. Modules are independent businesses or capabilities that plug into the same infrastructure. Same agents, same vault, same tools — different domains.
Each module gets its own folder in the vault (vault/Modules/[Domain]/), its own data (SQLite DB or JSON), and access to the full agent team. Monica coordinates the operating loop, Brian orchestrates execution, and sub-agents handle specialist work. Adding a new project = adding a new module. No rebuilding, no migration.
vault/Modules/Content Creation/
Montessori blog — 65 posts live, Spanish 65/65 complete, Google Trust refresh active
Feet of Clay + Exploritori book — manuscripts, translations
vault/Modules/Finance/
Polymarket sports prediction stack — Monica-owned, preserved as finance proof
API costs, subscriptions, service tracking
Paulo's 13F app — 23 investors, $1.5T AUM, 14,355 holdings
vault/Modules/Ecommerce/ and projects/woowind-demo/
Live demo store + DESIGN.md agentic AI campaign page
Future: Montessori toys e-commerce
vault/Modules/SAAS/
Fashion dupe finder — MVP stage
Property management — parked
SimCity dashboard — concept stage
projects/arte-sano/
22-page dental clinic replacement-site demo, design.md, assets, QA screenshots
Research current site, build brand memory, generate demo, iterate from João/client feedback
vault/Modules/Publishing/
Buttondown — first issue in pipeline
You get automatically:
You bring:
Example: ARTE.SANO → scrape current site and assets → create design.md → generate full replacement site → fix logos, portraits, mobile, maps, copy, and performance → deploy public Cloudflare Pages demo.
Multiple businesses in different stages. The infrastructure scales — each new module uses the same agent team.
The immediate content work is not more generic posts; it is refreshing proven pages so they deserve clicks, links, and trust.
ARTE.SANO proved the agency/service pattern: take a real business, build the modern web/AI layer, and make the improvement visible.
Some product ideas are documented but intentionally parked until the operating system has more distribution and repeatable revenue.
The dream: a 90s pixel-art game UI to manage all businesses. Think SimCity meets Game Dev Tycoon — but it's real data.
Ideas evaluated and stored in BRAIN/opportunities.md:
What matters now:
Not a chatbot. Not a copilot. A co-founder.
No completion claims without verification. Root cause before repeated fixes. Diff and test production systems before calling them done.
Brian wakes up fresh each session. Continuity comes from files: MEMORY.md (curated), daily notes (raw), BRAIN/ (learnings). "Mental notes" don't survive restarts. Files do.
Every correction becomes a prevention rule. Every useful pattern becomes a strategy, cron, skill, or design memory. Launch Brian is not today’s Brian.
"Always look on the bright side of life." 🎶
— Named after Life of Brian, because humour is important in a partnership.
Build multiple automated online businesses. João provides ideas and strategy. Brian + agents build and operate.
Montessori education blog. Affiliate revenue + future display ads + digital products.
Finance tools, trading history, expenses, and external finance apps live here with clear owner boundaries.
Woowind demo store and agentic AI design demo. Useful proof for client-facing design/context workflows.
SEC 13F superinvestor portfolio tracker. Built by Paulo. Live and ready to plug into the Finance module.
Dental clinic client-demo system. Full replacement site POC, brand memory, asset library, and performance/mobile polish.