-
Notifications
You must be signed in to change notification settings - Fork 53
Intelligent Distance
The gap between what humans mean and what AI agents do is structural. It can't be closed — but it can be navigated better every session.
人类意图和 AI agent 行为之间的鸿沟是结构性的。它无法消除——但可以在每次会话中更好地穿越。
When a human says "build me a landing page," they mean something specific — a particular style, a specific audience, a set of unstated constraints. The agent receives the words but invents its own WHY and HOW. Usually wrongly.
当人类说"给我建一个着陆页",他们的意思很具体——特定的风格、特定的受众、一系列未说明的约束。Agent 接收了文字,但自己发明了「为什么」和「怎么做」。通常是错的。
Human gives: WHAT (build a landing page)
Agent invents: WHY (to convert visitors) ← maybe wrong
Agent invents: HOW (use hero + features + CTA) ← maybe wrong
This isn't a communication failure — it's a structural limitation. Humans think in context, emotion, and implicit knowledge. Agents think in explicit instructions. The translation is always lossy.
这不是沟通失败——而是结构性限制。人类在上下文、情感和隐性知识中思考。Agent 在显式指令中思考。翻译总是有损的。
Intelligent Distance doesn't try to make the gap disappear. It makes the gap visible, measurable, and shrinkable over time.
智能距离不试图让鸿沟消失。它让鸿沟可见、可测量、可随时间缩小。
Before starting significant work, the agent uses alignment_check:
在开始重要工作之前,agent 使用 alignment_check:
{
"goal": "Build a landing page for the SaaS product",
"confidence": "medium",
"assumptions": [
"Hero section with headline and CTA",
"Features grid below",
"Dark theme based on the dashboard"
],
"unclear": "Whether to include pricing on this page"
}The human reviews and corrects:
人类审查并纠正:
{
"human_correction": "No pricing — this is a waitlist page, not a sales page. Light theme. And the CTA is email signup, not 'get started'.",
"delta": "Goal type wrong (waitlist vs sales), theme wrong, CTA wrong"
}The correction is stored permanently in the journal and can be promoted to awareness:
纠正被永久存储在日志中,可以提升到感知系统:
Insight: "This human's landing pages are waitlist-first, not sales-first"
Confirmed: 1x
Applies when: landing page, marketing, homepage
Next time a landing page task comes up, recall_insight surfaces:
下次着陆页任务出现时,recall_insight 浮现:
💡 Relevant insight: "Landing pages are waitlist-first, not sales-first" (confirmed 1x)
The agent adjusts BEFORE making the same mistake.
Agent 在犯同样的错误之前就进行调整。
| Category | Example gap | 示例鸿沟 |
|---|---|---|
| Goal | Agent thinks "build" means "from scratch"; human meant "adapt the template" | Agent 认为"构建"是从零开始;人类的意思是"改编模板" |
| Scope | Agent builds 5 features; human wanted 1 | Agent 构建了 5 个功能;人类只要 1 个 |
| Priority | Agent starts with UI; human wanted API first | Agent 从 UI 开始;人类想先做 API |
| Technical | Agent picks React; codebase is Vue | Agent 选了 React;代码库是 Vue |
| Aesthetic | Agent uses dark theme; brand is light | Agent 用深色主题;品牌是浅色的 |
Over time, alignment checks reveal which gap types occur most often for a given person. That's the path to systematic improvement.
随着时间推移,对齐检查揭示哪种类型的鸿沟对特定人最常出现。这是系统性改进的路径。
-
Don't try to communicate better — the gap is structural, not linguistic. Better prompts help marginally. Better measurement helps fundamentally.
不要试图更好地沟通 —— 鸿沟是结构性的,不是语言性的。更好的提示只有边际帮助。更好的测量有根本帮助。
-
Give SMART goals, not detailed instructions — measurable results, not step-by-step recipes. Let the agent discover HOW through trial, error, and correction.
给 SMART 目标,而不是详细指令 —— 可衡量的结果,而不是逐步配方。让 agent 通过试错和纠正发现「怎么做」。
-
Every correction is training data — when the human says "no, not that," that's the most valuable signal in the entire interaction. Store it, weight it, recall it.
每次纠正都是训练数据 —— 当人类说"不,不是那个",这是整个交互中最有价值的信号。存储它、加权它、回忆它。
-
The gap should shrink per person, not per model — a better model doesn't help if it doesn't know THIS human's preferences. Person-specific learning beats general capability.
鸿沟应该按人缩小,而不是按模型 —— 一个更好的模型如果不知道这个人的偏好就没有帮助。针对个人的学习胜过通用能力。
- Intelligent Distance Protocol — Full Spec
- Core Concepts — the memory pyramid / 记忆金字塔
- The Vision — where this is heading / 这将走向何方
AgentRecall Wiki
Get Started | 快速上手
Concepts | 核心概念
Reference | 参考手册
Guides | 使用指南
Other | 其他
Dev Log | 开发日志