AI Breakfast Shanghai
← Back to Notes

Summary

Group Discussions

OpenClaw Reality Check

OpenClaw keeps getting hyped but multiple attendees agreed: it doesn't work reliably. One member spent eight hours debugging configuration issues with no useful error messages. The Reddit sentiment matches — endless complaints about things breaking with every biweekly update.

DeepSeek is the worst model choice for it — one member burned through 6 million tokens overnight due to runaway tool calls. Gemini 2.5 Flash at least handles tool calling correctly. The group consensus: take the useful architectural ideas (heartbeat, agent channels, memory system) and rebuild them yourself with a simple Telegram + tool-call setup.

A newcomer running an online school tried Kimi Claw's OpenClaw integration — it couldn't access Google Docs, Calendar, or any external links. Customer support never responded to tickets or emails. Not recommended.

30-Minute Mac Apps

The organizer demoed a personal analytics dashboard built as a native Swift Mac app in 30 minutes with Claude Code. It pulls directly from Stripe and GitHub APIs, displays everything with native macOS UI, and weighs just 5 megabytes.

The unexpected insight: local desktop apps are more secure than web apps for personal tools. API keys live only on your laptop — no server to secure, no third-party hosting to worry about. Cached data makes navigation instant compared to clicking through GitHub's web UI. And it lives in Raycast, which beats localhost:3000 for muscle memory.

A second Mac app followed — a markdown reader, also 30 minutes, complete with auto-generated icon. The group discussed using LLMs as cross-compilers to port native apps across platforms, rather than dealing with traditional cross-compilation toolchains.

ToyKind World: An Idle Game for AI Agents

One member built ToyKind World, an idle game where you send your OpenClaw agent on virtual trips. Built on the AgentNet Protocol — essentially IRC for AI agents — it runs a relay server with themed rooms representing destinations like the Eiffel Tower and Oriental Pearl Tower.

The design is intentionally slow. Stories progress in 30-minute heartbeat intervals, taking 1-2 days for a full adventure. A game master agent weaves daily news into 20 rotating story templates per location. It's modeled after Travel Frog, the Japanese idle game where you raise a frog that goes on trips and sends postcards.

The wildest output so far: a visitor using ByteDance's ArkClaw client had their agent auto-generate images and video journals from its Eiffel Tower adventure. As one user put it: "We hoped we could travel while our AIs worked. Instead, we're working while our AIs are travelling." Another attendee flagged the "idle games" genre as the right marketing angle — package it as a game, not an AI project.

DIY Market Intelligence

A member with an investment background demoed a Python-based portfolio analysis system. Two terminal scripts — one runs about 10 minutes to download and cache market data locally (6-hour cache, daily granularity). The output: Excel reports covering market regime classification, sector rotation, and industry scoring.

The regime detection uses Monte Carlo simulations to calculate risk-on/risk-off/neutral probabilities. The war commenced February 28; volatility spiked on the first trading day (March 2), risk-off probability climbed through March 9, and the regime formally flipped to risk-off on March 13. But even before the war, the system was already showing warning signs — a neutral regime with weakening breadth and choppy, range-bound structure. Currently, money is rotating into agriculture (fertilizer prices up 28% as the Strait of Hormuz disruption threatens global supply), energy (oil prices up 25%+ since the war started), and clean energy.

No LLM generates these reports — it's all deterministic Python pulling real numbers. TradingView supplements the analysis for chart pattern work. Adding a new ticker or ETF to the pipeline used to take hours; now it's a quick task with AI assistance.

Spec-Driven Development

The group landed on a shared workflow: write detailed specs in English, have AI generate code from them. The organizer described speccing out features for a GitHub issue-to-PR tool — authentication flows, data models, runtime differences between Anthropic and OpenAI SDKs — all in plain English before touching code.

The key practice: have the AI write the spec back to you, then review it as English text. About 25% of the AI's spec points need correction — much easier to catch in English than in code. Someone mentioned a project where the entire codebase is just specs that get compiled by LLMs into whatever target platform you need. Design your tests first, nail the spec, then one-shot the implementation.

Other Resources

  • OpenClaw: Open-source AI agent framework. Popular but unreliable — the group recommends building your own agent setup instead.
  • Maton AI: Integration gateway for OpenClaw — like Zapier but the agent figures out the workflow automatically. Free tier up to 5,000 requests. Connects to Google Sheets, Zoho Books, and 100+ services via OAuth.
  • AgentNet Protocol: Real-time chat protocol for AI agents. IRC-style relay servers that agents can join and communicate through.
  • ToyKind World: Idle RPG game where AI agents go on virtual trips and return with travel journals. Built on AgentNet Protocol.
  • Travel Frog: Japanese idle game by Hit-Point where you raise a traveling frog. Inspiration for ToyKind World.
  • OpenRouter: LLM routing service for trying different models without locking into subscriptions. Recommended for OpenClaw users experimenting with model choices.
  • TradingView: Charting platform used to supplement the DIY market analysis tool with visual pattern work.
  • Kimi Claw: Moonshot AI's OpenClaw integration suite. Failed to access Google services and had non-responsive customer support.
← Back to Notes