Back to home

Compare

Comparing: I Talked to AI for 3 Weeks — It Got Dumber. One Prompt Fixed It & 我跟AI聊了三周,发现它越来越听不懂我 — 一个提示词救回来了

AEN
ClaudeAI对话技巧独立创业者·

I Talked to AI for 3 Weeks — It Got Dumber. One Prompt Fixed It

Scene: I thought AI would get smarter the more we talked. Wrong.

Last Wednesday afternoon, I was rewriting product copy in a coffee shop in Hangzhou. Same AI tool, twenty-plus rounds in — its replies kept getting shorter, more like it was brushing me off. I figured my prompts were the problem. Turns out this is a universal LLM flaw, not me.

What this is, and who's using it

Indie hacker levels.io published a prompt trick: at the start of every new chat, paste a "role-reset" instruction that forces the AI back into deep-thinking mode instead of sliding down an increasingly shallow context. While running his remote SaaS company SaaShip, he fires off hundreds of Claude chats daily — he says "this move saved my workflow."

Short version: stop relying on the AI to "remember context." Force it to reload your goals every time. My friend Xiaoyang, who runs a paid-knowledge course as a solo creator in Shenzhen, copied the method and finally got AI to stop "getting dumber the longer we talk."

Replication cost: $0, 5 minutes, zero technical skill

No paid tool. No code. Steps: open your usual AI chat (start a new conversation), grab levels.io's English prompt, translate it into Chinese for your business, then ask your actual question. First step: grab the original prompt here.

I made this mistake too — I assumed "the longer we chat, the better it knows me." It's actually the opposite. Longer conversations, more drift.

How to use this at your stage

Just starting, no clients yet: Skip for now. If you're only asking AI a few questions a day, you haven't hit context collapse yet. Bookmark this and come back when you're using AI heavily.

1-2 clients already: I'd push you to try this today. If you're editing copy with AI daily, you're right in the "context rot" sweet spot. Five minutes in could save you an afternoon of rewrites.

Scaling up, high content output: Turn that prompt into a template, auto-paste it into every new session. If you run courses, communities, or a personal IP, this is basically a must-have.

Last thing — this isn't magic, it's a habit tweak. But for anyone grinding with AI 2+ hours a day, the difference is real.

Source: levels.io
BZH
ClaudeAI对话技巧独立创业者·

我跟AI聊了三周,发现它越来越听不懂我 — 一个提示词救回来了

场景:我以为 AI 越聊越懂我,结果是相反的

上周三下午,我在杭州一家咖啡馆改产品文案。同一个 AI 工具,连续聊了二十多轮之后,它给的回复越来越短,越来越像在敷衍我。我以为是我提的问题太烂,后来才发现 — 这是 LLM 的通病,不是我的问题。

这方法是什么,谁在用

独立开发者 levels.io 公开了一个提示词技巧:每次开新对话时,先丢一段「角色重启」指令,让 AI 重新进入深度思考模式,而不是顺着越来越浅的上下文一路滑下去。他做远程公司 SaaShip 时,每天跟 Claude 对话上百条,他说「这招救了我的工作流」。

简单说,就是你不再依赖 AI 自己「记住上下文」,而是每次强制它重新加载你的目标。我做知识付费课程的朋友小杨,在深圳一个人运营 IP,照搬这套方法后,说 AI 终于不再「越聊越傻」了。

复刻成本:钱 0、时间 5 分钟、技术门槛零

不需要付费工具,不需要懂代码。具体步骤:打开你常用的 AI 对话框(新开一个会话),先复制粘贴 levels.io 那段英文 prompt,翻译成你业务的中文版本,然后正常问你的问题。第一步:点这个链接拿到原文 prompt。

我之前犯过一个错:以为「聊得久 AI 就更懂我」,其实是反的。对话越长,它越容易跑偏。

三类人怎么用

刚起步、还没客户:现在不试也没事。你一天可能就问 AI 几个问题,还没到上下文爆炸的程度。先收藏这篇文章,等你开始高强度用 AI 再回来。

已经有 1-2 个客户:我会强烈建议你立刻试。你每天跟 AI 改稿、对文案,这种强度正好是「上下文崩坏」的甜蜜点。5 分钟投入,可能省你一下午重写。

正在扩规模、内容产出量大:把这段 prompt 做成模板,每次开新会话自动粘贴。如果你做的是课程、社群、IP,这方法几乎算是刚需。

最后说一句:这不是什么神器,就是个习惯调整。但对你每天跟 AI 缠斗 2 小时以上的人,体感差别真的大。

Source: levels.io