不丢问题
目标和重要问题不会只留在一次聊天里。
从这里了解并开始使用 Inquiry Publishing Stack
专注于思考和创作，把记录、整理、检查和发布交给 AI 协助完成。
你不需要先学 GitHub、Agent、YAML 或 Cloudflare。AHICP 帮你把问题、资料、判断、决定和作品组织成可以长期继续的项目；AI 协助记录、整理、核对、执行和发布，关键判断仍由你确认。
阅读完整使用指南 我已经了解，直接开始1 分钟了解：为什么会需要这套体系
你可能会用 AI 学习、查资料、写作、研究、做项目。真正困难的通常不是“AI 能不能写一段文字”，而是几周、几个月以后，你还能不能知道：为什么这样做、证据在哪里、哪些只是猜测、哪些决定是你自己确认的、换一个 AI 后怎样继续。
你可以把注意力放在问题、判断和创作上；AI 协助记录、整理、核对和执行。重要的资料、依据和决定会被保存下来，让项目可以长期推进，也可以顺利交给新的 AI 继续。
不混证据
有来源支持的主张与证据、推断、建议和已经确认的决定彼此区分。
不怕换 AI
只要新的 AI / Agent 能读取项目状态，就可以从仓库恢复上下文，而不必依赖上一个模型的记忆。
关键决定由你确认
AI 可以承担很多工作，但重要判断、隐私边界和是否发布仍由你决定。
使用前后，最大的变化是什么
只有临时聊天
想到问题 → 问 AI → 得到答案 → 再问 → 几周后找不到依据 → 不知道哪些已经确认 → 换 AI 后重新解释。
项目有持续记录
提出问题 → 保存目标 → 保留来源 → AI 帮助分析 → 区分事实/推断/建议 → 你确认决定 → 保存当前状态 → 以后换人或换 AI 也能继续。
它带来的不只是效率，也是一套更清楚的工作方式
先把问题说清楚
先知道自己真正想解决什么，再急着要答案。
先看证据，再下判断
“资料说什么”和“我认为应该怎样”不是一回事。
允许明确说不知道
把未知和争议保留下来，比假装确定更可靠。
保存决定的原因
未来的你和新的 AI 都能理解为什么走到这里。
AI 建议不等于你的决定
重要提案经过你的确认以后，才成为项目正式采用的决定。
每次结束都留好下一步
结束时留下当前状态、下一步和仍待解决的问题，之后可以直接接着做。
为什么它能跨聊天、跨模型、隔很久以后继续
关键不在于“让 AI 记住一切”，而在于把不同类型的项目状态放在不同位置。聊天只是临时工作区，真正需要长期保留的内容写回项目仓库；新的 AI 接手时，再按固定顺序读取当前状态和相关长期记录。
Manifest + Context Interface
项目地图
告诉新的 AI：当前项目采用什么规则、哪些文件最重要、从哪里读取最新状态。它负责“找到信息”，而不是重复保存所有内容。
↓
Working Memory
当前正在做什么
Current Focus 保存当前最重要的目标、阻塞和下一步；Task Plan 保存正在进行的任务、待确认问题和计划；Work Log 留给需要回顾历史时使用，不要求每次都加载。
↓ 稳定结果写入长期记忆
Core + Decision Log
长期事实与决定
重要的人类决定进入 Decision Log，并更新相应的 Core。这样可以区分“AI 建议过什么”和“你真正确认过什么”。
↓
Framework
结构和论证
把核心内容组织成可以检查的结构。Working Framework 可以继续修改；只有经过确认的版本才成为 Approved Framework。
↓
Artifact + Evidence
作品与依据
文章、书、报告、网站等成果从上面的状态派生；证据和来源限制它能负责任地说什么。成果变化不能反过来偷偷改写上游决定。
仓库状态优先于聊天记忆重要状态写回仓库，换聊天或换 AI 不需要从头解释。
工作记忆保持简洁新 AI 先读当前焦点和任务计划，不需要把全部历史一次塞进上下文。
短期状态会“晋升”临时问题解决后，稳定结果进入 Decision Log、Core、Framework 或其他长期位置。
决定沿上游向下传播重要决定先记录，再更新结构，最后反映到作品，避免正文先变、依据却跟不上。
因此 AHICP 提供的不是“无限记忆”，而是一套可恢复、可追溯、能交接的项目记忆结构。
四个项目，各自帮你解决一个不同的问题
AHICP · 怎样让思考、资料和决定保持清楚？
帮助记录问题、证据、推理、决定、需要确认的事项、当前进度和交接信息。
帮助记录问题、证据、推理、决定、需要确认的事项、当前进度和交接信息。
PPF · 我的作品怎样长期保存、构建和发布？
管理权威源文件、构建、网页、版本、发布与 Continuous Web。
管理权威源文件、构建、网页、版本、发布与 Continuous Web。
Vault Interface · 怎样公开项目，又不暴露私人材料？
只公开经过确认的项目元数据，不要求公开私人研究、工作记录或私有仓库位置。
只公开经过确认的项目元数据，不要求公开私人研究、工作记录或私有仓库位置。
Starter · 我不懂技术，能不能让 AI 帮我把整套系统配好？
负责把各组件组合到项目中、选择配置、升级并维护机器可读设置。完整新项目默认使用 full AHICP + full PPF + Vault Interface。
负责把各组件组合到项目中、选择配置、升级并维护机器可读设置。完整新项目默认使用 full AHICP + full PPF + Vault Interface。
一个很简单的例子
假设一个学生想研究“为什么恐龙会灭绝”。他不需要先学习 Git。可以先让 AI 帮助保存问题和几份来源，把“已经有证据支持的事实”和“仍有争议的解释”分开，再由学生自己决定报告重点。最后如果要公开网页，只发布已经确认可以公开的最终内容，私人草稿仍然保持私有。
同一套方法也可以用于长期学习、写论文、写书、知识管理、课程项目或复杂创作。项目越长，“让未来还能理解和继续”就越重要。
AI 可以承担很多工作，关键决定仍由你确认
适合交给 AI 协助
找资料、整理、比较、总结、发现矛盾、起草、检查引用、维护文件、构建网页、执行已经授权的技术步骤，以及整理交接信息。
需要由你决定
为什么做这个项目、哪些问题值得问、是否接受重要解释、是否采用关键方案、哪些私人材料可以公开、是否正式发布，以及最终结果是否符合你的判断。
隐私和公开，是两个不同决定
源文件仓库可以一直保持 private（私有），同时提供 restricted（受限）网页；以后即使最终网页公开，源仓库也不必公开。原创、未发布和私人材料默认保持私有，密码、token 和私钥永远不应该写进仓库。
不懂技术也没关系：第一次只做五件事
告诉 AI 你想做什么。例如：“我想长期研究城市交通，最后写成一篇文章。”
告诉 AI 什么不能擅自决定。例如：“公开发布前必须先问我。”
要求使用完整默认配置。让 AI 按 full AHICP + full PPF + Vault Interface 建立项目，原创内容默认保持 private。
像平常一样工作。继续说“帮我找资料”“比较这两种解释”“记录我刚才的决定”。
让项目保存当前状态。结束重要工作时，让 AI 写回已经确认的决定、未解决问题和下一步。
你可以直接告诉 AI：
“请按照 Inquiry Publishing Stack 帮我建立这个长期项目。能自动完成的技术配置请尽量完成；只有需要我判断、批准、登录或直接输入秘密时再让我操作，并一步一步说明。”
“请按照 Inquiry Publishing Stack 帮我建立这个长期项目。能自动完成的技术配置请尽量完成；只有需要我判断、批准、登录或直接输入秘密时再让我操作，并一步一步说明。”
你先想清楚要做什么，技术配置交给 AI
从 AHICP 了解这套体系
先理解它为什么存在、怎样帮助长期思考与创作，以及四个组件各自做什么。
让 AI 从 Starter 开始配置
需要建立项目、恢复完整配置、选择方案、升级或部署时，由 AI 读取 Starter 的机器入口和相关技术契约。
完整使用指南
下面是完整指南，来自 build-info.json 所记录的版本化 AHICP 源码。站点在更新来源锁并重新部署后同步到新的维护版本；这里展示的是当前部署所固定的版本，你不需要前往 GitHub 才能阅读。
正在载入完整指南……
现在开始
如果你的 AI 能读取公开 URL 并具备所需工具，通常只需要一段启动指令。
开始项目不要求你先学会 GitHub、Starter、YAML、Cloudflare 或部署术语。接下来把启动指令发给 AI，由它进入机器入口并开始配置。
1
点击下面的“复制给 AI”。
系统会复制一段包含机器入口和工作规则的启动指令。
系统会复制一段包含机器入口和工作规则的启动指令。
2
打开你平常使用的 AI。
新建一个聊天，把刚才复制的内容粘贴进去。
新建一个聊天，把刚才复制的内容粘贴进去。
3
在最后加一句你真正想做的事。
例如：“我想长期研究城市交通，最后写成一本书。”然后发送。
例如：“我想长期研究城市交通，最后写成一本书。”然后发送。
4
接下来交给 AI 处理技术配置。
AI 应自行读取 https://inquirystack.philohub.workers.dev/agent/，恢复完整体系、建立或检查项目，并继续配置与验证。
AI 应自行读取 https://inquirystack.philohub.workers.dev/agent/，恢复完整体系、建立或检查项目，并继续配置与验证。
如果当前 AI 能访问公开 URL 并具备所需工具，发送以后通常不需要你自己打开机器入口或阅读技术文件。AI 应先确认已经读取入口，再确认你的项目目标，然后继续配置、检查和部署准备；如果能力或授权不足，应明确说明缺少什么并给出交接步骤。
Start here · Inquiry Publishing Stack
Stay focused on thinking and creating. Let AI help with the records, structure, checks, and publishing.
You do not need to learn GitHub, agents, YAML, or Cloudflare first. AHICP gives a long-running project a durable place for questions, sources, decisions, and work state, so AI can help without making the project depend on one chat or one model.
Read the complete guide I understand — start nowUnderstand it in one minute
You may use AI for learning, research, writing, or projects. The hard problem is often not whether AI can generate text. It is whether, weeks or months later, you can still tell why a decision was made, where the evidence came from, what was only a guess, what you actually approved, and how another AI can continue.
The point is not to make AI think for you. It is to keep the project understandable over time while AI helps with organization, checking, execution, and other technical work.
Keep the question
Goals and important questions do not disappear into one conversation.
Keep evidence separate
Source-supported claims and evidence, inferences, suggestions, and approved decisions remain distinguishable.
Change AI without starting over
If a new AI or agent can read the project state, it can reconstruct context from the repository instead of depending on the previous model's memory.
Keep important decisions explicit
AI can do a great deal of work, while important judgments, privacy choices, and publication decisions stay explicit.
What changes when the project keeps its own state
Temporary chat only
Have a question → ask AI → receive an answer → ask again → weeks later lose the evidence → cannot tell what was confirmed → explain everything again to a new AI.
Durable project state
Ask a question → preserve the goal → retain sources → AI helps analyze → separate fact/inference/suggestion → human confirms decisions → save state → another human or AI continues.
It supports a clearer way of working, not just faster output
Clarify the question
Understand what you are really trying to solve before rushing to an answer.
Evidence before judgment
What a source says and what you conclude are not the same thing.
Preserve uncertainty
Explicitly recording “unknown” is often more reliable than pretending certainty.
Preserve reasons
Your future self and a new AI can understand why the project reached its current state.
An AI proposal is not a project decision
A proposal becomes part of the project only after it has been reviewed and accepted.
Leave a clear next step
End each work session with the current state, next actions, and unresolved questions easy to find.
Why the project can survive across chats, models, and long gaps
AHICP does not try to make a model remember everything. It separates short-lived operational state from durable project memory, then gives a new AI a predictable route through both.
Manifest + Context Interface
Project map
Points a new AI to the rules, canonical files, and current state. It tells the agent where to look instead of duplicating the project everywhere.
↓
Working Memory
What is happening now
Current Focus holds the highest-priority objective, blocker, and next action. Task Plan holds active work, pending decisions, clarifications, and next steps. Work Log is available for retrospective history, but is not loaded by default.
↓ stable results are promoted
Core + Decision Log
Durable commitments
Important accepted decisions are recorded and reconciled with the relevant Core, keeping a visible boundary between AI proposals and decisions the project has actually adopted.
↓
Framework
Structure and reasoning
The project’s claims and structure are organized into an inspectable framework. A Working Framework can change; an Approved Framework becomes a stable baseline only after review.
↓
Artifact + Evidence
Outputs and support
Articles, books, reports, and sites are derived from the upstream state. Evidence constrains what can responsibly be claimed, and downstream edits do not silently rewrite upstream decisions.
Repository state outranks chat memoryImportant state is written back to the project, so a new chat or model does not require a full restart.
Working Memory stays smallA new AI starts with current focus and active tasks instead of loading the entire history.
Short-term state gets promotedOnce an issue is resolved, durable results move into the appropriate long-term location.
Decisions propagate from upstream to downstreamRecord the decision first, update the structure next, and only then update the artifact.
The goal is not infinite context. It is recoverable, traceable project memory that another AI can pick up reliably.
Four projects, each solving a different part of the problem
AHICP · How do I keep questions, evidence, and decisions clear over time?
Organizes inquiry, evidence, reasoning, decisions, current work state, and handoff.
Organizes inquiry, evidence, reasoning, decisions, current work state, and handoff.
PPF · How does my work remain durable, buildable, and publishable?
Governs canonical source, builds, web output, versions, publication, and Continuous Web.
Governs canonical source, builds, web output, versions, publication, and Continuous Web.
Vault Interface · How can a project be public without exposing private work?
Supplies approved public metadata without requiring private research, working memory, or private repository locators.
Supplies approved public metadata without requiring private research, working memory, or private repository locators.
Starter · I do not know the technical details. Can an AI compose this correctly?
Handles composition, adoption, profiles, upgrades, and machine-readable configuration. Complete new projects default to full AHICP + full PPF + Vault Interface.
Handles composition, adoption, profiles, upgrades, and machine-readable configuration. Complete new projects default to full AHICP + full PPF + Vault Interface.
A simple example
Imagine a student studying why dinosaurs went extinct. The student does not need to learn Git first. An AI can help preserve the question and sources, separate well-supported evidence from disputed explanations, and maintain the project structure. The student still chooses the focus. If the final work becomes a website, approved output can be public while private drafts remain private.
The same pattern scales to long-term learning, papers, books, knowledge management, courses, and complex creative work.
AI can take on a lot of work; important decisions still need your judgment
Good work to delegate to AI
Finding sources, organizing, comparing, summarizing, detecting inconsistencies, drafting, citation checks, repository maintenance, website builds, authorized technical execution, and handoffs.
Decisions that still need your judgment
Why the project exists, which questions matter, whether to accept important interpretations, which proposals to adopt, what private material may become public, whether to publish, and responsibility for the final result.
Privacy and publication are different decisions
You may keep the source repository private while operating a restricted Web publication. Even if the final website becomes public later, the source repository does not need to become public. Original and unpublished work stays private by default, and secrets never belong in repositories.
No technical background? Do only five things first
Tell the AI what you want to do. “I want to study urban transportation for several months and eventually write an article.”
Say what the AI must not decide alone. “Ask me before public publication.”
Request the complete default stack. Ask for full AHICP + PPF + Vault Interface, keeping original work private by default.
Work normally. Ask for sources, comparisons, explanations, and recording of your decisions.
Make the project remember. At the end of important work, ask the agent to write confirmed decisions, unresolved questions, and next steps back into project state.
“Set up this long-term project according to the Inquiry Publishing Stack. Complete as much technical configuration as you can. Ask me only when I need to judge, approve, sign in, or enter a secret, and give me clear step-by-step instructions.”
Learn the system here; let the AI enter the technical setup
Use AHICP to understand the system
Start with why the system exists, how it supports long-running work, and what each component is for.
Let the AI enter through Starter
For project setup, reconstruction, upgrades, and deployment, the AI should use Starter’s machine entry and retrieval contract.
Complete guide
The full guide below comes from the versioned AHICP source recorded in build-info.json. The site moves to a newer maintained revision when its source lock is refreshed and redeployed; this page therefore shows the revision pinned by the current deployment.
Loading the complete guide…
Start now
If your AI can read public URLs and has the required tools, one bootstrap instruction is usually enough.
You do not need to learn GitHub, Starter, YAML, Cloudflare, or deployment terminology before you begin. Send the instruction below to your AI; it should enter the machine setup and continue from there.
1
Click “Copy for AI” below.
This copies the machine entrypoint and the operating rules.
This copies the machine entrypoint and the operating rules.
2
Open the AI you normally use.
Start a new conversation and paste the copied instruction.
Start a new conversation and paste the copied instruction.
3
Add one sentence describing what you actually want to do.
For example: “I want to study urban transportation over the long term and eventually write a book.” Then send it.
For example: “I want to study urban transportation over the long term and eventually write a book.” Then send it.
4
Let the AI handle the technical setup.
It should read https://inquirystack.philohub.workers.dev/agent/, reconstruct the full stack, create or inspect the project, and begin configuration and verification.
It should read https://inquirystack.philohub.workers.dev/agent/, reconstruct the full stack, create or inspect the project, and begin configuration and verification.
If the AI can access public URLs and has the required tools, you normally should not need to open the machine entry or read technical files yourself. It should confirm that it loaded the entry, confirm your project goal, and continue the engineering setup; if capability or authorization is missing, it should state what is missing and provide a handoff path.