Watch on YouTube: https://www.youtube.com/watch?v=wBzfjrkTKvM
Channel: Coco哥
How One Creator Automates Trading Research, Content, and Apps With AI Agents
Coco哥’s latest walkthrough shows what a daily AI-driven workflow looks like when you stop treating large language models as chatbots and start deploying them as persistent agents. The video covers a stack that includes Claude Max, ChatGPT Max, Grok, and a custom OpenClaw agent running on Cloudflare Workers — all coordinating to produce market research, publish a financial content site, build a mobile game, edit videos, and even flag phishing emails in real time. The creator also highlights a practical constraint: Hong Kong users currently lack direct access to Claude and ChatGPT’s top tiers, pushing the conversation toward viable alternatives.
What the Video Covers
- A fleet of AI agents handling news aggregation, market-close reports, a Telegram deep-dive bot, a stock “thermometer,” and a backtesting engine hooked to Interactive Brokers for auto-trading.
- A deep backtest of the “January Effect” strategy (buy the prior year’s S&P 500 winner each January) with full drawdown and recovery analysis.
- A complete rebuild of
cashmate.hkusing Claude Code on Cloudflare Workers — auto-generating daily hot topics, weekly reports with charts, a 10-year YouTube knowledge base (200+ summaries), a course system, a market monitor indicator, and a credit-card miles calculator. - End-to-end video production automation (subtitles, animations, music, Reels editing) running roughly 90% hands-free.
- A Mahjong game (“跑馬仔”) built via Grok and ChatGPT o1 with zero manual code, deployed to both App Store and Play Store.
- Real-time scam detection on a fake police email, plus an affiliate pitch for Surfshark One.
- Cost breakdown: ~£180/month for Claude Max + ChatGPT Max, with weekly token limits hit from automated content generation.
Step-by-Step Walkthrough
1. Agent Infrastructure: From Home Server to Cloud
The journey started in February 2024 with OpenClaw running on a local Linux box. Today the entire stack lives on Cloudflare Workers — no local hardware required. Agents pull X (Twitter) feeds, summarize into structured analyses, push to cashmate.hk, and trigger Telegram alerts. The market-close report agent runs after each session; the “thermometer” agent scores sector breadth daily. The backtesting engine connects to Interactive Brokers’ API, allowing strategy validation before any capital is committed.
2. January Effect Backtest: The Full Research Cycle
The agent was tasked with testing a simple rule: each January 1, buy the previous year’s top-performing S&P 500 stock. From 2012–2025 the hypothetical $100 grew to $170,000 (1,700×) versus the S&P 500’s 8×. The agent then stress-tested variations — top 3, top 5, top 10 holdings; dollar-cost averaging $10K/year — and produced interactive charts. Crucially, it flagged that SanDisk’s 6× surge in the final year (2024–25) inflated the headline number, and that the strategy suffered an 84% drawdown (2005–2011) and a 97% max drawdown (down to $3 from $100), requiring eight years to recover. The output was a complete research report, not just a number.
3. Rebuilding cashmate.hk With Claude Code
Using Claude Code on Cloudflare Workers, the creator prompted a full rewrite: server code, frontend design, database schema, and deployment pipeline. The site now auto-generates five daily hot-topic analyses with charts, a weekly market report, a searchable 10-year YouTube knowledge base (200+ video summaries), a course delivery system, a market monitor (7 of 8 historical red signals preceded crashes; currently flashing red), and a miles calculator for credit-card points. The creator emphasizes that AI cannot design intent — every feature, user flow, and data source had to be specified explicitly.
4. Zero-Code Mobile Game: Mahjong “跑馬仔”
Grok produced a playable prototype; ChatGPT o1 iterated through dozens of bug fixes (rule edge cases, scoring, multiplayer sync). The creator wrote no code. The build was submitted to both iOS and Android stores. This illustrates the shift from “AI writes snippets” to “AI delivers shippable products” when the human acts as product manager and QA.
5. Video Pipeline & Scam Detection
Subtitles, B-roll animations, background music, and vertical Reels cuts are now 90% automated. In a live demo, a suspicious “police” email was forwarded to an agent; it identified spoofed headers, mismatched domains, and social-engineering language within seconds.
Tools, Numbers, and Costs
- Claude Max: £90/month
- ChatGPT Max: £90/month
- January Effect backtest (2012–2025): $100 → $170,000 (1,700×) vs. S&P 500 $100 → $800 (8×)
- SanDisk contribution: 6× in final year (2024–25)
- Max drawdown: 97% (peak to trough); 84% during 2005–2011
- Recovery period: 8 years to break even from 2005 peak
- Market monitor: 7/8 historical red signals preceded crashes; 3 most recent each −20%; current signal ~10–30 days before average crash
- Surfshark One promo: code
COCOGOR= 4 extra months, 30-day refund - Tech stack: Cloudflare Workers, Interactive Brokers API, Grok, ChatGPT o1, Claude Code, OpenClaw
- YouTube knowledge base: 200+ summaries spanning 10 years
Caveats & Hype Check
The 1,700× return is real backtest data but cherry-picked: the start date (2012) sits at a market bottom, and the final year’s result is dominated by a single stock (SanDisk). A 97% max drawdown makes the strategy practically untradeable for any meaningful capital. The creator explicitly states the market monitor signal is historical analysis, not a prediction, and takes no positions based on it. The Surfshark mention is a disclosed affiliate sponsorship. No personal trading P&L from the auto-trading system is claimed.
Additional constraints: Hong Kong users cannot directly subscribe to Claude Max or ChatGPT Max. The creator suggests Gemini, Copilot, Kimi, and DeepSeek as functional alternatives. Token limits are hit weekly due to the volume of automated content generation on cashmate.hk.
Concrete Next Actions
- Visit
cashmate.hkfor free daily market analysis, weekly reports, and the 10-year YouTube knowledge base. - Test the “跑馬仔” Mahjong game (iOS/Android) and submit feedback.
- Try Surfshark One with code
COCOGORfor data-leak monitoring and masked email (30-day refund). - If you’re in Hong Kong, evaluate Gemini, Copilot, Kimi, and DeepSeek as Claude/ChatGPT alternatives.
- Define a repetitive research or operational task, then deploy an AI agent to run it for hours or days unattended.
- Backtest any trading idea with an agent before deploying capital — verify drawdowns, recovery periods, and single-stock concentration.
- Automate a content pipeline: YouTube → transcripts → summaries → knowledge base → course modules.
- Forward suspicious emails or messages to an AI agent for real-time scam analysis.
Leave a Reply