AI UNDERDOGSDAILY PICK
AI UNDERDOGS
别信 AI 说「完成」
Stop trusting AI's 'done'
LeoStehlik/proof-loop
你让 AI 写完代码，它说搞定了
但你真敢直接跑吗？
Proof Loop 的思路很简单——不信任
只验证
Your coding agent says 'done.' But do you
actually trust that?
Proof Loop flips the script: don't trust, verify
验证者 ≠ 建造者
Verifier ≠ Builder
核心改动：写代码的 AI 和检查代码的 AI
不能是同一个
Proof Loop 给每个任务配一个独立验证者角色
专门挑毛病
这直接瞄准了 AI agent 最大的软肋——自己给自己验收
The key move: the agent that writes code
and the agent that verifies it are two
separate roles. Proof Loop pairs every task with
a dedicated verifier. This hits AI agents' core
weakness — self-verification
★ SIGNAL 1
不是嘴上说说
Not just vibes
每个任务都定义了明确的验收标准
agent 不能光说「我写完了」
得交出 proof artifacts——结构化的证据
你能直接翻看验证
不是 prompt 工程的话术
是实实在在的东西
Every task has explicit acceptance criteria. The agent
can't just wave its hands — it must
produce proof artifacts: structured evidence you can actually
inspect. Not prompt-engineered hand-waving, real deliverables
★ SIGNAL 2
验收标准写进协议
Criteria baked into protocol
很多人做 agent 框架
验证那一步就是个摆设
Proof Loop 不一样——验收标准是任务定义的一部分
验证者拿到的是原封不动的 criteria
不是二次翻译的 prompt
这才是真正能跑起来的闭环
Most agent frameworks treat verification as a checkbox
Here, acceptance criteria are part of the task
definition itself — the verifier gets the original
criteria, not a re-prompted summary. That's how you
build a loop that actually closes
作者原话说得很直白：「I make my coding agents prove they finished the task
」没有吹什么革命性框架
就是从自己被坑的经历出发
做了一个朴素但有效的机制
The author's own words: 'I make my coding
agents prove they finished the task.' No revolutionary
claims — just a straightforward mechanism born from
getting burned one too many times
AI UNDERDOGS
让它证明给你看
Make it prove it
LeoStehlik/proof-loop
关注 · 每天发现更多 AI 神作
github.com/LeoStehlik/proof-loop