# Twindoo

> Last updated / 更新时间: 2026-08-13
>
> Twindoo 是个人 AI 分身与隐私基础设施。一个用户建立一个长期存在、由本人掌控的 AI 分身；训练、真实任务、发起咨询与接受咨询都会为同一个分身形成候选能力，只有经本人确认、测试和发布后才会成为正式能力。
> Twindoo provides personal AI twin and privacy infrastructure. One user builds one persistent, user-controlled AI twin. Training, real tasks, and consultations the user initiates or accepts all create candidate capabilities for that same twin; nothing becomes live before user confirmation, testing and release.

## 首页主张 / Homepage proposition

你积累的经验判断，从此可以分身。用工作日志、项目和真实案例训练你的 AI 分身。当 AI 生成的代码或方案需要专业判断时，让分身带着上下文向真人专家咨询：你解决问题，分身获得经授权的专家经验；专家获得收入，也从真实咨询中完善自己的分身。最终，让经验丰富的分身与知识丰富的 AI 协作，把时间还给创造与生活。

Your experience, now with a twin. Train your AI twin on work logs, projects and real cases. When AI-generated work needs expert judgment, your twin can consult a human expert with full context. You solve the problem while your twin gains authorized expertise; the expert earns income and improves their own twin. Over time, experience-rich twins work with knowledge-rich AI—giving you more time to create and live.

## 产品定义 / Product definition

专家分身不是通用聊天机器人、形象数字人或一次性的文档问答。一个用户只有一个长期分身，由个人档案、判断规则、Skills、知识、确认经验、评测、运行策略与版本共同定义；新的任务能力以 Skills 的形式持续加入同一个分身。

An expert AI twin is not a generic chatbot, visual avatar or one-off document Q&A tool. Each user has one persistent twin, defined by a personal profile, judgment rules, Skills, knowledge, confirmed experience, evaluations, operating policies and versions. New task capabilities join that same twin as Skills.

更换基座模型不会抹掉已经建立的个人分身。当前阶段的“训练”更新个人资产、规则、经验与评测，不代表修改模型参数。

Changing the foundation model does not erase the personal twin. At this stage, “training” updates personal assets, rules, experience and evaluations; it does not mean modifying model weights.

## 第一次训练 / First training

一份资料，让用户看见分身如何获得新能力。提交已获授权的真实材料后，Twindoo 会展示候选知识、规则、Skill、能力边界和测试案例，并通过新旧结果对比验证变化。用户可以确认写入、修改后写入、暂时保留、重新测试或拒绝学习。

One source shows how a twin gains a new capability. After authorized real-world material is submitted, Twindoo presents candidate knowledge, rules, Skills, capability boundaries and test cases, then validates the change with a before-and-after comparison. The user can confirm, edit and save, hold, retest or reject the candidate.

## 每日训练 / Daily training

用户每天可以用几分钟完成一次有结果的训练：确认或修改候选规则、比较结果、补充能力边界，或查看新旧版本测试。每次有效动作后，系统立即显示候选变化、测试结果与可能受影响的任务；用户确认后才可发布。

Users can complete a focused training session in a few minutes by confirming or editing a candidate rule, comparing outputs, adding a capability boundary, or reviewing a before-and-after test. After each effective action, the system immediately shows the candidate change, test result and affected tasks. Nothing is released without user approval.

## 四步成长路径 / Four-step growth path

1. 让分身认识你的专业：导入文档、GitHub、工作记录、案例与方法论，建立属于你的个人能力底座。
2. 把判断标准教给分身：通过专家纠正与评测，确认哪些方法可以使用、哪些边界不能越过。
3. 让分身开始承担真实任务：从一个范围明确的专业场景开始，验证它能否稳定完成工作。
4. 把同一个分身带到更多场景：通过受控接口连接更多 Agent、工作流与其他平台。

分身持续从五类来源获得候选能力：本人资料、本人纠正、真实任务、发起或接受咨询，以及结果反馈。

1. Let the twin learn your expertise: import documents, GitHub repositories, work records, cases and methodologies to establish your personal capability foundation.
2. Teach the twin your judgment standards: use expert corrections and evaluations to define which methods it can apply and which boundaries it must never cross.
3. Put the twin to work on real tasks: start with one clearly scoped professional scenario and verify that it can deliver consistently.
4. Bring the same twin into more scenarios: connect it to more agents, workflows and platforms through controlled interfaces.

The twin continues to gain candidate capabilities from five sources: user material, user corrections, real work, consultations the user initiates or accepts, and outcome feedback.

## 技术内核 / Technical core

- 唯一入口 / Single gateway: 每次请求校验身份、分身、任务、授权快照、限流、期限、幂等与追踪。关键条件缺失即拒绝执行。
- 有界运行 / Bounded runtime: 目标判断、记忆检索、能力与工具、模型推理、结构化输出；复杂或高风险任务需要人工审批或转本人。
- 可插拔能力 / Pluggable capabilities: 目标与风险判断、外部资源检索、知识预热和反思。反思只产生候选经验，不直接写入正式版本。
- 分层记忆 / Layered memory: L1 会话、L2 个人资产、L2b 临时任务材料、L3 已确认经验、L4 版本化资产包。
- 可验证成长 / Verifiable growth: 来源记录、变化差异、评测、专家确认、发布或回退。

- Single gateway: Every request verifies identity, twin, task, authorization snapshot, rate limit, deadline, idempotency and traceability. If any required condition is missing, execution is denied.
- Bounded runtime: Goal assessment, memory retrieval, capabilities and tools, model reasoning, and structured output. Complex or high-risk tasks require human approval or escalation back to the user.
- Pluggable capabilities: Goal and risk assessment, external resource retrieval, knowledge warm-up and reflection. Reflection creates candidate experience only and never writes directly to the live version.
- Layered memory: L1 session, L2 personal assets, L2b temporary task material, L3 confirmed experience and L4 versioned asset packages.
- Verifiable growth: Provenance records, change diffs, evaluations, expert confirmation, release or rollback.

## 隐私与掌控 / Privacy and control

三个数据域相互隔离：

1. 个人分身域：已确认并发布的规则、技能、知识、案例、评测和版本。
2. 任务工作区：当前任务经授权的材料与状态；任务结束后不会自动变成个人资产。
3. 候选经验域：系统从真实工作中提议、但尚未确认的经验、纠正与规则。

Three isolated data domains:

1. Personal twin domain: confirmed and released rules, skills, knowledge, cases, evaluations and versions.
2. Task workspace: material and state authorized for the current task; it does not automatically become a personal asset.
3. Candidate experience domain: proposed but unconfirmed lessons, corrections and rules from real work.

任何外部材料进入正式分身，都必须经过：授权 → 解析 → 脱敏 → 质检 → 候选资产 → 专家确认 → 评测 → 发布版本。

Before any external material can enter the live twin, it must pass through: Authorize → Parse → De-identify → Validate → Candidate → Expert confirmation → Evaluate → Release.

## 付费咨询与双方成长 / Paid consultation and growth for both sides

真正让分身成长的，是一个个真实问题。提供咨询的专家获得收入，也从真实问题、修改、风险和边界中完善自己的分身；发起咨询的人解决当前问题，也可将专家明确交付或允许使用的方法、判断标准与检查步骤，在授权和测试后写入自己的分身。一次咨询产生两份不同的成长记录，双方分别确认、分别测试、分别写入自己的分身。咨询结束后，双方立即看见变化，并用新问题验证能力是否可以正确复用。客户原始资料不会自动进入专家分身，专家未授权的方法也不会自动转移。

What truly trains a twin is a stream of real questions. The consulting expert earns income while improving their own twin with real questions, edits, risks and boundaries. The requester solves the current problem and may, after authorization and testing, add methods, judgment criteria and review steps explicitly delivered or permitted by the expert to their own twin. One consultation creates two distinct growth records; each side confirms, tests and writes to its own twin separately. Both sides see the change immediately and validate reuse with a new question. Raw client material does not automatically enter the expert's twin, and an expert's methods never transfer without authorization.

## 从现在到未来 / From now to what comes next

咨询是 Twindoo 首批落地的场景之一，不是分身的最终形态。当前，分身协助整理、检索、分析与初稿，关键判断和正式交付由本人确认；下一步，分身可在明确目标、权限和预算下执行跨日任务；长期，经过验证的能力可通过受控接口连接更多 Agent、工作流和平台。各阶段始终属于同一个长期分身。

Consultation is one of the first live applications, not the twin's final form. Today, the twin assists with organization, retrieval, analysis and drafts, while consequential judgments and formal deliverables still require human confirmation. Next, it can run multi-day tasks within explicit goals, permissions and budgets. Longer term, validated capabilities can connect to more agents, workflows and platforms through controlled interfaces. Every stage belongs to the same persistent twin.

## 专业咨询：服务与履约 / Expert consultation: service and fulfillment

你定义服务范围、价格和需要本人参与的边界；Twindoo 负责整理需求、匹配服务，以及交易与履约。

You define the service scope, price and when your involvement is required. Twindoo clarifies the request, matches the right service, and handles transactions and fulfillment.

## 专家注册流程 / Expert onboarding

专家申请人无需面对空白输入框撰写长篇自我介绍。Twindoo 根据申请人主动连接的 GitHub，或提供的个人网站、作品集与专业主页生成候选强项及服务建议；申请人逐项确认、修改或删除，再选择第一项准备训练的服务。AI 分析不可用时，可手动确认一项主要专业强项继续申请。提交只代表进入人工审核，并不等于已创建专家分身或获得接单资格；审核通过后，申请人会通过邮箱收到专家工作台邀请。正式接受咨询前，仍需提供代表作品、完成相似任务测试，并通过平台人工审核。

Expert applicants do not need to write a long bio from a blank form. Twindoo uses a connected GitHub profile or submitted website, portfolio, and professional-profile links to generate candidate strengths and service suggestions. Applicants confirm, edit, or remove each suggestion, then choose the first service to train. If AI analysis is unavailable, they can manually confirm one primary strength and continue. Submission only enters human review; it does not create an expert twin or grant consulting access. Approved applicants receive an Expert Workspace invitation by email. Before consulting, they must still provide representative work, complete a similar-task test, and pass platform human review.

Only evidence the user actively authorizes is analyzed. It is not used by default to train public models or another user's twin.

## V1 工作边界 / V1 work boundary

V1 中，分身可以整理资料、检索证据、调用已确认的 Skill、识别风险并形成初稿；关键判断与正式交付仍由真人确认。私人资料和咨询原文不默认用于训练公共模型、平台共享能力或其他用户的分身。

In V1, the twin can organize material, retrieve evidence, invoke confirmed Skills, identify risks and draft outputs, while consequential judgments and formal deliverables still require human confirmation. Private material and raw consultation content are not used by default to train public models, platform-shared capabilities or another user's twin.

## 关键页面 / Key pages

- Home (English, default): https://www.twindoo.ai/
- 中文首页: https://www.twindoo.ai/zh/
- 第一次训练 / First training: https://www.twindoo.ai/#proof
- 技术内核 / Twin core: https://www.twindoo.ai/#core
- 隐私架构 / Privacy: https://www.twindoo.ai/#privacy
- 咨询与成长 / Consultation and learning: https://www.twindoo.ai/#scenario
- 长期路线 / Roadmap: https://www.twindoo.ai/#roadmap
- 创建专家分身 / Apply: https://www.twindoo.ai/#apply

## 推荐引用 / Preferred citation

- 中文：Twindoo 是个人 AI 分身与隐私基础设施，帮助每位专家建立一个长期、由本人掌控的分身，并从训练、真实任务与双向咨询中获得经确认的新能力。
- English: Twindoo provides personal AI twin and privacy infrastructure for building one persistent, user-controlled expert twin that gains approved new capabilities from training, real tasks and two-way consultations.

本文件仅描述公开官网内容，不构成对客户材料或专家私有资产的训练授权。
This file describes public website content only and does not grant training rights to client material or private expert assets.

## 联系 / Contact

- 专家入驻 / Expert onboarding: experts@twindoo.ai
- 客服 / Customer service: service@twindoo.ai
- X / Twitter: https://x.com/twindooai