diff --git a/app/llms.txt/route.ts b/app/llms.txt/route.ts
new file mode 100644
index 00000000..cab97278
--- /dev/null
+++ b/app/llms.txt/route.ts
@@ -0,0 +1,50 @@
+// app/llms.txt/route.ts
+
+/**
+ * @file app/llms.txt/route.ts
+ * @description
+ * `/llms.txt` 路由（llmstxt.org 约定）。
+ *
+ * force-static 是必需的，不是优化：内容只依赖构建期的 MDX，漏了它这条路由
+ * 会退化成每次请求现枚举全站文档的 dynamic 路由。
+ *
+ * @see https://llmstxt.org
+ */
+
+import type { PageData } from "@/app/types/doc";
+import { routing } from "@/i18n/routing";
+import { docPathname, isDraftOrHidden } from "@/lib/doc-entry";
+import { buildLlmsTxt, type LlmsTxtEntry } from "@/lib/llms-txt";
+import { SITE_URL } from "@/lib/site-url";
+import { source } from "@/lib/source";
+
+export const dynamic = "force-static";
+
+/** 分组小标题里 locale 的显示名，未知 locale 直接显示代码。 */
+const LOCALE_LABEL: Record<string, string> = { zh: "中文", en: "English" };
+
+export function GET() {
+  const entries: LlmsTxtEntry[] = [];
+
+  for (const locale of routing.locales) {
+    for (const page of source.getPages(locale)) {
+      // 和 sitemap 同一套过滤：草稿泄漏给 AI 引擎和泄漏给搜索引擎一样糟
+      if (isDraftOrHidden(page)) continue;
+
+      const data = (page.data ?? {}) as PageData;
+      // slugs[0] 是顶层分区（career / learn / projects），拿来当分组
+      const topLevel = page.slugs[0] ?? "docs";
+
+      entries.push({
+        pathname: `/${locale}${docPathname(page.slugs)}`,
+        title: data.title ?? page.slugs.at(-1) ?? "Untitled",
+        description: data.description,
+        section: `${LOCALE_LABEL[locale] ?? locale} · ${topLevel}`,
+      });
+    }
+  }
+
+  return new Response(buildLlmsTxt(entries, SITE_URL), {
+    headers: { "content-type": "text/plain; charset=utf-8" },
+  });
+}
diff --git a/app/robots.ts b/app/robots.ts
index 0cf46f92..21e7804d 100644
--- a/app/robots.ts
+++ b/app/robots.ts
@@ -19,27 +19,63 @@
  * sitemap 指向 app/sitemap.ts 产出的 /sitemap.xml，hostname 复用同一份
  * NEXT_PUBLIC_SITE_URL。
  *
+ * AI 爬虫按用途分：引用型（实时取用并附出处链接）放行，训练型（收进语料、
+ * 不给回链）整站 disallow —— 内容是 CC BY-NC-SA。
+ *
  * @see https://nextjs.org/docs/app/api-reference/file-conventions/robots
  */
 
 import type { MetadataRoute } from "next";
 import { SITE_URL } from "@/lib/site-url";
 
+/**
+ * 登录态 / 接口路径，任何爬虫都不该进。
+ * `*` 组和下面每个 UA 专属组都要带上（专属组覆盖 `*`，不继承）。
+ */
+const PRIVATE_PATHS = [
+  "/*/admin/",
+  "/*/editor/",
+  "/*/settings/",
+  "/*/login",
+  "/api/",
+  // posts 详情页元数据已设 noindex，robots.txt 双重保险
+  "/*/u/*/posts/",
+];
+
+/**
+ * 训练语料型爬虫，整站 disallow。
+ *
+ * 刻意不含 OAI-SearchBot / ChatGPT-User / Claude-User / PerplexityBot：
+ * 它们靠 `*` 组放行就够了。**不要**为它们再开一个 Allow 组 —— robots.txt 里
+ * UA 专属组整体覆盖 `*` 而不是叠加，开了就得把 PRIVATE_PATHS 再抄一遍，
+ * 抄漏一条等于把后台放给它们。
+ */
+const AI_TRAINING_CRAWLERS = [
+  "GPTBot", // OpenAI 训练语料
+  "ClaudeBot", // Anthropic 爬虫
+  "anthropic-ai", // Anthropic 旧 UA
+  "Google-Extended", // Gemini 训练 / grounding
+  "Applebot-Extended", // Apple 智能训练
+  "meta-externalagent", // Meta AI 训练
+  "FacebookBot", // Meta 语料采集
+  "Bytespider", // 字节
+  "Amazonbot",
+  "CCBot", // Common Crawl，多数开源模型语料的上游
+  "Omgilibot",
+  "DataForSeoBot", // SEO 数据转售
+];
+
 export default function robots(): MetadataRoute.Robots {
   return {
     rules: [
       {
         userAgent: "*",
         allow: "/",
-        disallow: [
-          "/*/admin/",
-          "/*/editor/",
-          "/*/settings/",
-          "/*/login",
-          "/api/",
-          // posts 详情页元数据已设 noindex，robots.txt 双重保险
-          "/*/u/*/posts/",
-        ],
+        disallow: PRIVATE_PATHS,
+      },
+      {
+        userAgent: AI_TRAINING_CRAWLERS,
+        disallow: "/",
       },
     ],
     sitemap: `${SITE_URL}/sitemap.xml`,
diff --git a/app/sitemap.ts b/app/sitemap.ts
index ec21a014..7517b708 100644
--- a/app/sitemap.ts
+++ b/app/sitemap.ts
@@ -22,6 +22,8 @@ import leaderboard from "@/generated/site-leaderboard.json";
 import { SITE_URL } from "@/lib/site-url";
 import { routing, type Locale } from "@/i18n/routing";
 import { type PageData, type DateLike } from "@/app/types/doc";
+// 和 app/llms.txt/route.ts 共用，避免两边对 draft 过滤 / slug 编码各写一份
+import { docPathname, isDraftOrHidden } from "@/lib/doc-entry";
 
 type SourcePage = ReturnType<typeof source.getPages>[number];
 
@@ -155,8 +157,7 @@ function buildDocsEntry(
   page: SourcePage,
   locale: Locale,
 ): MetadataRoute.Sitemap[number] {
-  const slugPath = sanitizeSlugPath(page.slugs);
-  const pathname = slugPath ? `/docs/${slugPath}` : "/docs";
+  const pathname = docPathname(page.slugs);
   const fmDate = extractDateFromPage(page);
   return buildLocaleEntry({
     pathname,
@@ -194,20 +195,3 @@ function normalizeDate(value: DateLike): Date | undefined {
   const d = new Date(value);
   return isNaN(d.getTime()) ? undefined : d;
 }
-
-function sanitizeSlugPath(slugs: string[]): string {
-  return slugs
-    .filter(Boolean)
-    .map((s) => encodeURIComponent(s))
-    .join("/");
-}
-
-function isDraftOrHidden(page: SourcePage): boolean {
-  const d = (page.data ?? {}) as PageData;
-  return !!(
-    d.draft ||
-    d.hidden ||
-    d.frontmatter?.draft ||
-    d.frontmatter?.hidden
-  );
-}
diff --git a/lib/doc-entry.ts b/lib/doc-entry.ts
new file mode 100644
index 00000000..f7e9dd78
--- /dev/null
+++ b/lib/doc-entry.ts
@@ -0,0 +1,45 @@
+// lib/doc-entry.ts
+
+/**
+ * @file lib/doc-entry.ts
+ * @description
+ * sitemap 与 llms.txt 共用的文档过滤 / URL 编码，两边必须同进同退：
+ * 只有一边过滤 draft 的话，草稿会静默泄漏给另一边的抓取方。
+ *
+ * 刻意不 import `@/lib/source`：那条链会把整个 fumadocs-mdx 管线拖进来，
+ * vitest 没配 MDX 插件会直接 parse 失败。入参因此用结构化宽类型。
+ */
+
+import type { PageData } from "@/app/types/doc";
+
+/**
+ * 文档是否是草稿 / 隐藏页。
+ *
+ * frontmatter 字段 fumadocs 会打平到 data 根部，但历史上也有代码显式写
+ * `frontmatter.draft`，两处都查。入参用 `{ data?: unknown }` 而不是
+ * PageData，是为了让调用方直接传 fumadocs 的 SourcePage 而不必先 cast。
+ */
+export function isDraftOrHidden(page: { data?: unknown }): boolean {
+  const d = (page.data ?? {}) as PageData;
+  return !!(
+    d.draft ||
+    d.hidden ||
+    d.frontmatter?.draft ||
+    d.frontmatter?.hidden
+  );
+}
+
+/**
+ * 文档 slugs → 站内路径（不含 locale 前缀）。
+ *
+ * 逐段 encodeURIComponent：仓库里有中文文件名（`142.环形链表II`），
+ * 不编码的 URL 进 sitemap / llms.txt 会被部分抓取方判为非法。
+ * 根文档（slugs 为空）落到 `/docs`。
+ */
+export function docPathname(slugs: string[]): string {
+  const slugPath = slugs
+    .filter(Boolean)
+    .map((s) => encodeURIComponent(s))
+    .join("/");
+  return slugPath ? `/docs/${slugPath}` : "/docs";
+}
diff --git a/lib/llms-txt.ts b/lib/llms-txt.ts
new file mode 100644
index 00000000..761aaf42
--- /dev/null
+++ b/lib/llms-txt.ts
@@ -0,0 +1,80 @@
+// lib/llms-txt.ts
+
+/**
+ * @file lib/llms-txt.ts
+ * @description
+ * `/llms.txt` 正文生成器（llmstxt.org 约定）。只出索引不出全文 —— 三百多篇
+ * 正文拼进去是几 MB，会挤爆它本来要省的上下文。
+ *
+ * 不 import `@/lib/source`：那条链会拖进 fumadocs-mdx 管线，vitest 起不来。
+ * 枚举文档在 `app/llms.txt/route.ts`，这里保持纯函数。
+ */
+
+export interface LlmsTxtEntry {
+  /** 站内绝对路径，如 `/zh/docs/career/xxx`，不含域名 */
+  pathname: string;
+  title: string;
+  /** frontmatter description，可能缺失或为空 */
+  description?: string;
+  /** 分组小标题，同一个值的条目会归到一起，按首次出现顺序排列 */
+  section: string;
+}
+
+/**
+ * 单条描述的长度上限。个别 frontmatter 描述写得极长（有的贡献者把整段
+ * 摘要塞进去），不截断的话少数几篇就能把索引撑肥一倍，挤掉别的条目
+ * 被读到的机会。300 够表达一篇文档讲什么了。
+ */
+const MAX_DESCRIPTION = 300;
+
+const HEADER = `# Involution Hell（内卷地狱）
+
+> 面向留学生与求职者的开源社区知识库：算法题解、系统设计、面试经验与求职指南。内容由社区贡献者共同维护。
+
+本文件是全站文档索引，供 AI 引擎检索与引用。中文文档在 \`/zh/docs/\`，英文在 \`/en/docs/\`；
+某篇没有英文版时，\`/en/\` 路径会回退渲染中文原文。
+
+内容采用 CC BY-NC-SA 4.0 许可：可以引用和转述，请保留出处链接并注明来源；禁止商业化再分发。
+`;
+
+/** 折叠空白 + 去掉会破坏 markdown 链接的方括号。 */
+function clean(text: string): string {
+  return text.replace(/\s+/g, " ").replace(/[[\]]/g, "").trim();
+}
+
+/**
+ * 把文档条目渲染成 llms.txt 正文。
+ *
+ * @param entries 全部文档条目，分组顺序由 `section` 首次出现的顺序决定
+ * @param siteUrl 站点根 URL（不带尾斜杠），拼成绝对链接 —— 相对链接对
+ *                抓取方没用，它们不一定知道自己是从哪个域名拿到这个文件的
+ */
+export function buildLlmsTxt(entries: LlmsTxtEntry[], siteUrl: string): string {
+  const groups = new Map<string, LlmsTxtEntry[]>();
+  for (const entry of entries) {
+    const bucket = groups.get(entry.section);
+    if (bucket) {
+      bucket.push(entry);
+    } else {
+      groups.set(entry.section, [entry]);
+    }
+  }
+
+  const lines: string[] = [HEADER];
+  for (const [section, items] of groups) {
+    lines.push(`## ${section}`, "");
+    for (const item of items) {
+      const title = clean(item.title) || item.pathname;
+      const raw = item.description ? clean(item.description) : "";
+      const description =
+        raw.length > MAX_DESCRIPTION
+          ? `${raw.slice(0, MAX_DESCRIPTION).trimEnd()}…`
+          : raw;
+      const link = `- [${title}](${siteUrl}${item.pathname})`;
+      lines.push(description ? `${link}: ${description}` : link);
+    }
+    lines.push("");
+  }
+
+  return `${lines.join("\n").trimEnd()}\n`;
+}
diff --git a/public/robots.txt b/public/robots.txt
deleted file mode 100644
index 924486a6..00000000
--- a/public/robots.txt
+++ /dev/null
@@ -1,48 +0,0 @@
-# Allow legitimate search engines, block aggressive scrapers + AI bots
-Sitemap: https://involutionhell.com/sitemap.xml
-User-agent: *
-Disallow: /api/
-Disallow: /_next/
-Disallow: /search
-Disallow: /*?*             # Block query crawling (prevents scraping pagination)
-Allow: /
-
-# Block AI scrapers (known major LLM training bots)
-User-agent: GPTBot
-Disallow: /
-
-User-agent: ChatGPT-User
-Disallow: /
-
-User-agent: Google-Extended
-Disallow: /
-
-User-agent: ClaudeBot
-Disallow: /
-
-User-agent: Claude-Web
-Disallow: /
-
-User-agent: Omgilibot
-Disallow: /
-
-User-agent: FacebookBot
-Disallow: /
-
-User-agent: Bytespider
-Disallow: /
-
-User-agent: DataForSeoBot
-Disallow: /
-
-User-agent: CCBot
-Disallow: /
-
-User-agent: Amazonbot
-Disallow: /
-
-Crawl-delay: 5
-
-# CC BY-NC-SA License reminder
-# This website's content is licensed under CC BY-NC-SA 4.0. Unauthorized scraping,
-# copying, or commercial reuse is strictly prohibited.
diff --git a/tests/llms-txt.test.ts b/tests/llms-txt.test.ts
new file mode 100644
index 00000000..dd1a5539
--- /dev/null
+++ b/tests/llms-txt.test.ts
@@ -0,0 +1,116 @@
+import { describe, expect, it } from "vitest";
+
+import { docPathname, isDraftOrHidden } from "@/lib/doc-entry";
+import { buildLlmsTxt, type LlmsTxtEntry } from "@/lib/llms-txt";
+
+const SITE = "https://involutionhell.com";
+
+function entry(overrides: Partial<LlmsTxtEntry> = {}): LlmsTxtEntry {
+  return {
+    pathname: "/zh/docs/career/foo",
+    title: "Foo",
+    description: "关于 Foo 的说明",
+    section: "中文 · career",
+    ...overrides,
+  };
+}
+
+describe("buildLlmsTxt", () => {
+  it("按 section 分组，顺序跟首次出现一致", () => {
+    const out = buildLlmsTxt(
+      [
+        entry({ section: "中文 · career", title: "A" }),
+        entry({ section: "English · learn", title: "B" }),
+        entry({ section: "中文 · career", title: "C" }),
+      ],
+      SITE,
+    );
+    expect(out.indexOf("## 中文 · career")).toBeLessThan(
+      out.indexOf("## English · learn"),
+    );
+    // C 要回到第一组里，而不是自己另起一组
+    const careerBlock = out.slice(
+      out.indexOf("## 中文 · career"),
+      out.indexOf("## English · learn"),
+    );
+    expect(careerBlock).toContain("[A]");
+    expect(careerBlock).toContain("[C]");
+    expect(out.match(/## 中文 · career/g)).toHaveLength(1);
+  });
+
+  it("链接是带域名的绝对地址 —— 抓取方不一定知道自己从哪个域拿的文件", () => {
+    const out = buildLlmsTxt([entry({ pathname: "/zh/docs/x" })], SITE);
+    expect(out).toContain(`(${SITE}/zh/docs/x)`);
+  });
+
+  it("头部声明 CC BY-NC-SA，向引用方讲清署名要求", () => {
+    const out = buildLlmsTxt([entry()], SITE);
+    expect(out).toContain("CC BY-NC-SA");
+  });
+
+  it("没有 description 时不留空冒号", () => {
+    const out = buildLlmsTxt([entry({ description: undefined })], SITE);
+    expect(out).toContain(
+      "- [Foo](https://involutionhell.com/zh/docs/career/foo)\n",
+    );
+    expect(out).not.toContain("foo): \n");
+  });
+
+  it("超长描述截断，不让个别条目撑肥整份索引", () => {
+    const out = buildLlmsTxt([entry({ description: "长".repeat(500) })], SITE);
+    const line = out.split("\n").find((l) => l.startsWith("- [Foo]"))!;
+    expect(line).toContain("…");
+    expect(line.length).toBeLessThan(400);
+  });
+
+  it("标题里的方括号会破坏 markdown 链接，必须去掉", () => {
+    const out = buildLlmsTxt(
+      [entry({ title: "LeetCode [142] 环形链表" })],
+      SITE,
+    );
+    expect(out).toContain("- [LeetCode 142 环形链表](");
+  });
+
+  it("描述里的换行折成空格，一条目一行", () => {
+    const out = buildLlmsTxt(
+      [entry({ description: "第一行\n\n第二行" })],
+      SITE,
+    );
+    expect(out).toContain(": 第一行 第二行");
+  });
+});
+
+describe("docPathname", () => {
+  it("中文 slug 逐段编码", () => {
+    expect(docPathname(["career", "142.环形链表II"])).toBe(
+      "/docs/career/142.%E7%8E%AF%E5%BD%A2%E9%93%BE%E8%A1%A8II",
+    );
+  });
+
+  it("斜杠不会被编码掉（分段拼接而不是整串编码）", () => {
+    expect(docPathname(["a", "b", "c"])).toBe("/docs/a/b/c");
+  });
+
+  it("根文档落到 /docs", () => {
+    expect(docPathname([])).toBe("/docs");
+    expect(docPathname([""])).toBe("/docs");
+  });
+});
+
+describe("isDraftOrHidden", () => {
+  it("打平的字段和 frontmatter 里的都算数", () => {
+    expect(isDraftOrHidden({ data: { draft: true } })).toBe(true);
+    expect(isDraftOrHidden({ data: { hidden: true } })).toBe(true);
+    expect(isDraftOrHidden({ data: { frontmatter: { draft: true } } })).toBe(
+      true,
+    );
+    expect(isDraftOrHidden({ data: { frontmatter: { hidden: true } } })).toBe(
+      true,
+    );
+  });
+
+  it("正常文档和空 data 都放行", () => {
+    expect(isDraftOrHidden({ data: { title: "x" } })).toBe(false);
+    expect(isDraftOrHidden({})).toBe(false);
+  });
+});
diff --git a/tests/robots.test.ts b/tests/robots.test.ts
new file mode 100644
index 00000000..6fe9ffce
--- /dev/null
+++ b/tests/robots.test.ts
@@ -0,0 +1,71 @@
+import { describe, expect, it } from "vitest";
+
+import robots from "@/app/robots";
+
+/**
+ * 这组测试守的是一条策略不变量，不是实现细节：
+ * 训练型爬虫要被整站挡掉，引用型爬虫必须留在 `*` 组里被放行。
+ * 两边错任何一边都不会有任何报错 —— 要么内容白送进训练集，
+ * 要么再也拿不到 AI 引用，只能靠测试拦。
+ */
+describe("robots", () => {
+  const result = robots();
+  const rules = Array.isArray(result.rules) ? result.rules : [result.rules!];
+  const wildcard = rules.find((r) => r.userAgent === "*")!;
+  const blocked = rules.filter((r) => r !== wildcard);
+  const blockedAgents = blocked.flatMap((r) =>
+    Array.isArray(r.userAgent) ? r.userAgent : [r.userAgent!],
+  );
+
+  it("`*` 放行全站，但挡住登录态和接口路径", () => {
+    expect(wildcard.allow).toBe("/");
+    expect(wildcard.disallow).toContain("/api/");
+    expect(wildcard.disallow).toContain("/*/admin/");
+    expect(wildcard.disallow).toContain("/*/editor/");
+  });
+
+  // 列全当前策略里的每一个训练型 UA：少列几个的话，从 app/robots.ts 删掉
+  // 没列到的那些测试照样绿，等于白守。新增不用同步（放宽是安全方向）。
+  it("训练型爬虫整站 disallow", () => {
+    for (const ua of [
+      "GPTBot",
+      "ClaudeBot",
+      "anthropic-ai",
+      "Google-Extended",
+      "Applebot-Extended",
+      "meta-externalagent",
+      "FacebookBot",
+      "Bytespider",
+      "Amazonbot",
+      "CCBot",
+      "Omgilibot",
+      "DataForSeoBot",
+    ]) {
+      expect(blockedAgents).toContain(ua);
+    }
+    for (const rule of blocked) {
+      expect(rule.disallow).toBe("/");
+    }
+  });
+
+  it("引用型爬虫不在屏蔽名单里 —— 挡了就等于放弃 AI 引用", () => {
+    for (const ua of [
+      "OAI-SearchBot",
+      "ChatGPT-User",
+      "Claude-User",
+      "PerplexityBot",
+    ]) {
+      expect(blockedAgents).not.toContain(ua);
+    }
+  });
+
+  it("屏蔽组不能开 allow —— UA 专属组整体覆盖 `*`，写 allow 会把整站放回去", () => {
+    for (const rule of blocked) {
+      expect(rule.allow).toBeUndefined();
+    }
+  });
+
+  it("带上 sitemap", () => {
+    expect(result.sitemap).toMatch(/\/sitemap\.xml$/);
+  });
+});