给 AI 交易代理的第一条需求,应该是“未经批准,不准碰真钱”。
这条 Minara 长提示词把任务拆得很具体:先询问仓位、日亏损、持仓数量和停机条件;再要求列出数据来源,把进出场与失效条件写清楚。回测要计入手续费和滑点,还得换一段没参与调参的数据复测。
其中一条很实用:连跳过的交易也要记日志,说明触发了什么、为什么没做。最后才是模拟运行,实盘另等人工批准。
可以学它怎样给 Agent 写任务边界和验收条件。这是一份提示词模板,不能证明平台已把这些约束做成硬限制,更不能把回测数字当预期收益。
Replying to @noisyb0y1
Paste this into Minara:
I want a trading bot that runs 24/7, and I want to see every piece of it before anything goes live. Research and backtest first, paper trading second, real money only when I say so.
Start with my limits, not with a strategy. Ask me for: max position size, max daily loss, max open positions, which assets I will touch, and the one condition that shuts everything down. Write them down and treat them as fixed. If a setup breaks any of them, skip the trade, never resize around it.
Then research the market I named. Pull live prices, on-chain flows, fundamentals and sentiment, and show me the sources for everything you use. Tell me which signals have actually predicted moves in this asset and which ones are just noise that looks convincing.
Build the strategy as something I can read: entry condition, exit condition, invalidation, position size, and which signals have to agree before it fires. If the signals disagree, the default is no trade.
Backtest it on minute-level history across the longest period available. Report number of trades, win rate, average win and loss, worst drawdown, and what fees and slippage take out. Then re-run it on a period you did not tune on and show me both results side by side.
Be honest about the gap. Tell me how much of the edge survives real execution at my size. If the edge is thinner than the cost of trading it, say so and stop there.
Run it on paper for a defined period before anything real. Log every entry with the reason attached: what fired, how strongly, what would invalidate it, and what you were unsure about. Log the trades you skipped too.
When it goes live, keep the same log. Every morning give me one short report: what it did overnight, what it skipped and why, current exposure, and anything that drifted from the backtest.
Never change my limits. Never place a trade that breaks them. Never tell me a backtested number is expected profit. If something looks wrong, stop and ask instead of guessing.
End every report with: What it did, What it skipped, What would break this.
Esto es una puta locura. Andrej Karpathy trabajó 8 años en OpenAI y Tesla.
La semana pasada condensó todo lo aprendido en una clase gratuita de 2 horas.
Agents → Harness → Loops → Graphs
Hay bootcamps de 15.000 dólares que enseñan menos que esto.
Puede que ahora no tengas 2 horas.
Guárdalo y míralo antes de que se pierda en tu feed.
绝了,AI 修图的需求里,“哪里不能变”终于写清楚了。
这组10条 ChatGPT 提示词模板,把修饰、换背景、换衣服、扩图拆开讲。拿换背景这条,直接改成自己的场景就能用:
“把背景换成[新场景]。保留人物的脸、姿势、衣服、发型和比例;让光线方向、阴影、反射、景深、大小与透视匹配新环境。”
删杂物也要写明补回后面的纹理,扩图则指定画幅,并要求保留原图区域。这些是可修改的提示词模板,成图还得检查细节。把要改的、要保留的和衔接标准一起写进需求,才方便判断哪一步需要重做。
绝了,给 SaaS 做产品演示视频,这套工具链把素材交接都安排好了。
这段20秒图解,把流程排成:找参考、用 Excalidraw 画分镜、用 Playwright 截产品界面,再交给 Claude 配合 Remotion 做动画,补音效、编码导出。字体用 Fontsource,图标用 Lucide,少一些临时拼凑。
起步命令也给了:在视频项目里运行 npx skills add remotion-dev/skills,把官方制作规范交给编程代理。Mobbin 是付费参考库,Freesound 音效要逐条看授权。
可以拿来试做功能讲解和上新视频。一个产品、一个功能、一条讲清楚的故事,已经是很具体的交付范围。


