Midmap 是专为与智能体协作而设计的信息空间。在设计 Midmap 的过程中,我对用户和智能体各自的角色与定位做了许多取舍。从面向对话协作,转变为面向画板协作,让每一轮讨论都有交代,每一次修改都可巡检。
12
8
626
分享下我目前规划开发需求的方式:
1. 我提思路,让 Agent 给草案,并按主题将信息铺开。
2. 我看着内容做澄清,审内容,提要求。Agent 按我说的分别更新。
3. 澄清结束后,让智能体看着完整信息评估是否可开发,我做答疑。
4. 答疑结束后,让智能体设计开发方案和验收标准,我来审核。
5. 全部审阅结束后,进入开发。
这个过程周期很长,但同时解决了两个问题:
1. 这个过程首先会让我把整个需求理清楚,讲明白。需求不是许愿,是一次响应用户具体诉求的程序化方案。
2. 信息不在一开始澄清和定调,大概率会在 Agent 犯错后被动澄清,而犯错可能是有代价的。
提前澄清,快不如慢。
3
5
53
Managing time and energy while vibe coding is a new challenge too.
People often complain about having to switch between and supervise multiple sessions. But honestly, the urge to catch problems early and correct them keeps pulling at my attention.
If an agent can make mistakes while writing code, I can’t fully trust an agent’s automated code review either.
So I think we first need a way for agents to correct themselves during development. One approach is TDD: write the tests first, then use them to check the implementation as it develops. That might help catch things going off track along the way.
Self-correction can’t solve everything, though. TDD helps check whether the code produces the expected behavior. It can’t necessarily catch problems with how things are implemented, such as architecture, performance, or other side effects.
So reviewing the work afterward is still necessary. Until I can comfortably let go, I need to find a balance between where I put my attention and how I spend my time.
Give the agent some time to work on its own, and first make sure that when it’s time to check the homework, there’s actually homework to check.
That way, I can free up a continuous block of time and energy for other things, like scrolling through X.
2
1
66
在 Vibe Coding 过程中的开发时间和精力管理也是新课题。
多 Session 的切换监督确实被人诟病,但坦白说,及时发现异常并尽早纠正的念头会持续拉扯注意力。
如果 Agent 自动写代码的过程中已经可能发生问题,那 Agent 的 自动 Code Review 也是无法被信任的。
所以,需要先构建 Agent 在开发过程中的自纠错机制。比如借助 TDD 思路,让测试先行,通过开发过程与测试用例的碰撞来做纠错。过程中途可能发生的方向偏离就可能被纠正。
自纠错无法解决所有问题,TDD 是对预期行为结果上的纠偏。还无法解决实现方式的隐患,比如架构、性能、以及可能引入的其它副作用。
所以后检查依然必要,在无法撒手之前,只能先平衡注意力和时间控制。
给 Agent 一段自由活动时间,优先确保收作业的时候交上来的是份作业。
这样,可以腾挪一段连续的时间和精力出来处理其它事情,比如刷会儿 X。
1
47
Writing requirements for an AI agent:
Day 1: Make a wish. (You can do anything, right?)
Day 2: More details. (So you don’t know everything.)
Day 3: Don’t change that! (Please stop breaking things.)
Day 4: Do as I say! (Help me, don’t make a mess.)
Day 5: Can you write it? (I’m tired of writing requirements.)
Day 6: I’ll do it. (Your requirements got me in trouble.)
Day 7: God, help me.
1
1
49
What's interesting to me is not "more agents."
It's that control flow is leaving the chat. A plan that lives only in the conversation drifts. A script you can read, approve, and — in Antigravity — save and rerun is a different kind of plan.
The model still does the work. I want to review the program that directs it.
3
55
Product Hunt launch is coming to an end soon! If you haven’t checked out Midmap yet, I’d really appreciate your support. Thanks so much!
Midmap: A shared information space for working with your AI agent
producthunt.com/products/mid… via @ProductHunt
3
5
124
写文写了一个多小时,现在已经是 Midmap 上线 Product Hunt 的第四个小时了。当前评论席冷清,名次潜力也很大。如有朋友前来支持,感激不尽:
producthunt.com/products/mid… View article
1
108














