Spent the last few months building an AI prompt system for DeFi/altcoin research, and the biggest lesson: generic "analyze this token" prompts are useless. GPT-4o (and most models) will happily hallucinate confident-sounding fundamentals if you don't structure the request.

What actually works:

1. Force a framework, not a vibe check. Give the model a fixed rubric — team/tokenomics/unlock schedule/on-chain activity/protocol risk — and make it fill in each section or say "unknown." Otherwise it skips straight to a narrative.

2. Separate on-chain interpretation from narrative interpretation. Wallet flows, liquidity depth, and holder concentration are a different analysis task than "what's the market saying on Twitter." Mixing them in one prompt muddies both.

3. Make risk assessment its own step. Contract risk, unlock cliffs, and protocol exploit history need a dedicated pass — not a footnote at the end of a bullish writeup.

4. Journal every trade thesis, not just the entries. A DCA prompt that logs why you bought at each level is worth more than a P&L screenshot six months later when you're trying to figure out what your edge actually was.

Built this into a full bundle (research/on-chain/risk/thread-gen/journal prompts) if anyone wants to skip the trial-and-error — but even just the 4 principles above will upgrade whatever you're already doing.