diff --git a/CLAUDE.md b/CLAUDE.md
index 2b1158d..749cc44 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -17,7 +17,7 @@ The backend was rewritten from Django in 2026-10 to cut memory use (~25 MB vs. 1
   - `src/auth.ts` — Django-compatible password hashing; JWT login, no cookies: `POST /account/login` returns a 15-min HS256 `access_token` (sent as `Authorization: Bearer`, signed with `JWT_SECRET`, random per process if unset) and a 14-day opaque `refresh_token` (only its sha256 is stored, in `auth_refresh_token`; `POST /account/refresh` trades it for a new access token, logout / password change / deactivation delete it). An invalid or expired Bearer token is a 401 even on public endpoints so the client knows to refresh. `requireUser/requireAdmin/requireSuper` (401 / 403)
   - `src/routes/` — `account`, `tasks` (tutorial / challenge / creation), `files` (task assets + image upload), `submission`, `showcase` (mounted under `/submission/showcase`), `review` (mounted under `/submission/review`: teacher class × task wall — no `classname` means all classes —, students' peer works with `scope=class|all`), `prompt`
   - Who can read a submission (`canView` in `routes/submission.ts`): staff, the owner, anyone if it's in the showcase, or any student who has submitted the same task (any class). Students' `GET /submission/` only lists their own. Peer works and the random-rating item are anonymous until the viewer rates them (`PUT /:id/score` returns the author). `flag` now only carries `present` (teacher's in-class presentation queue); the old four colors are no longer shown
-  - `src/services/` — `score.ts` (rating weights super×0.5, admin×0.3, normal×0.2, then Bayesian smoothing with C=3 and zones), `gradebook.ts`, `llm.ts` (OpenAI-compatible streaming via fetch; DeepSeek default, ARK for doubao models; chat context is the last 20 messages with older code blocks elided), `classifier.ts` (Bloom-level L1–L6 prompt tagging, runs in background), `cascade.ts` (Django's CASCADE deletes done by hand, plus the shared delete rules: `hasRatingsOrAwards`, `deleteRound`), `pyround.ts` (`pyRound`/`pySum` reproduce Python's `round()` and `sum()` so scores match the old system bit-for-bit)
+  - `src/services/` — `score.ts` (rating weights super×0.5, admin×0.3, normal×0.2, then Bayesian smoothing with C=3 and zones), `gradebook.ts`, `llm.ts` (OpenAI-compatible streaming via fetch; DeepSeek default, ARK for doubao models; chat context is the last 20 messages with older code blocks elided), `classifier.ts` (Bloom-level L1–L6 prompt tagging, runs in background; the model sees the task brief and the previous prompt, temperature 0, prompts under 4 chars get no level), `cascade.ts` (Django's CASCADE deletes done by hand, plus the shared delete rules: `hasRatingsOrAwards`, `deleteRound`), `pyround.ts` (`pyRound`/`pySum` reproduce Python's `round()` and `sum()` so scores match the old system bit-for-bit)
   - `src/ws.ts` — `/ws/prompt/<task_id>/?token=<access_token>` (code generation chat; the AI context and the page's chat include `conversation`, `edit` and `manual` messages) and `/ws/guidance/<task_id>/?token=...` (prompt coaching); one message per connection at a time (a message sent while the AI is still replying gets an error, payloads capped at 64 KB); on disconnect the pending user message is deleted and the LLM request aborted
   - `scripts/` — one-off CLIs: `create-superuser.ts`, `classify-prompts.ts`, `recalculate-scores.ts`
   - Roles: `super`, `admin`, `normal`. Task model is multi-table: `task_task` + one of `task_tutorial` / `task_challenge` / `task_creation` (`task_ptr_id`)
diff --git a/client/src/utils/helper.ts b/client/src/utils/helper.ts
index b0fd0d2..599f853 100644
--- a/client/src/utils/helper.ts
+++ b/client/src/utils/helper.ts
@@ -27,12 +27,12 @@ export function sourceLabel(source: string | null | undefined) {
 
 // 提问的 Bloom 等级（services/classifier.ts 后台打的标签）
 export const PROMPT_LEVELS: Record<number, { name: string; desc: string }> = {
-  1: { name: "记忆", desc: "复述或识别知识点" },
-  2: { name: "理解", desc: "用自己的话解释概念" },
-  3: { name: "应用", desc: "把知识用到新情境" },
-  4: { name: "分析", desc: "拆解结构、找出规律" },
-  5: { name: "评价", desc: "按标准作出判断" },
-  6: { name: "创造", desc: "整合信息做出新东西" },
+  1: { name: "记忆", desc: "只说要做什么，没有自己的要求" },
+  2: { name: "理解", desc: "说出看到的现象或疑问" },
+  3: { name: "应用", desc: "提出具体要求（颜色、布局、功能），照搬题目要求也算" },
+  4: { name: "分析", desc: "说清在什么操作下出了什么问题，或推测原因" },
+  5: { name: "评价", desc: "按标准判断好坏并说明理由，或比较不同做法" },
+  6: { name: "创造", desc: "题目之外自己的设计：主题、玩法、交互" },
 }
 
 export function levelLabel(level: number | null | undefined) {
diff --git a/server/src/services/classifier.ts b/server/src/services/classifier.ts
index a313b21..0ff1e0c 100644
--- a/server/src/services/classifier.ts
+++ b/server/src/services/classifier.ts
@@ -1,19 +1,27 @@
 import { all, get, run } from "../db"
 import { complete, llmConfigured } from "./llm"
 
-const CLASSIFY_SYSTEM_PROMPT = `你是一个教育评估专家。根据布鲁姆认知分类学，分析以下学生在前端学习中发送给AI助手的一条提示词，判断该提示词所体现的认知层级。
+// 学生在这里是“指挥 AI 做网页”，看的是这句提问里学生自己想了多少，不是要 AI 做的东西有多大：
+// 把题目要求原样贴过去，功能列得再多也只是照着做（L3）
+const CLASSIFY_SYSTEM_PROMPT = `你是一个教育评估专家。学生在网页制作课上让 AI 生成或修改网页，下面给出题目要求、学生上一轮的提问和这一轮的提问。请按布鲁姆认知分类学，判断“这一轮提问”体现的学生自己的思考层级。
 
 层级定义：
-- L1 记忆：能背诵HTML标签语法（例："帮我写一个按钮"）
-- L2 理解：能解释flex布局原理（例："为什么这里不居中？"）
-- L3 应用：能独立搭建页面结构（例："用flex做导航栏，间距16px"）
-- L4 分析：能定位跨浏览器兼容性bug（例："Safari中margin失效，原因？"）
-- L5 评价：能对比并选择方案（例："对比Grid与Flex方案优劣"）
-- L6 创造：能设计并实现原创交互作品（例："设计夜间/日间切换效果"）
+- L1 记忆：只说要做什么，没有自己的要求，或者原样复述题目名称（例："帮我做一个象棋游戏"、"继续"）
+- L2 理解：用自己的话说出看到的现象或疑问，但没有具体要求（例："为什么点了没反应"、"还是没有显示"）
+- L3 应用：给出具体、可执行的要求，如颜色、布局、要加的功能；照搬或改写题目要求里的功能清单也属于这一级（例："方块大一点，旁边加一个橡皮擦和清空按钮"）
+- L4 分析：说清楚在什么操作下出了什么问题，或推测原因、拆解问题出在哪部分（例："拖动绘制时松开鼠标还在画，是不是没监听 mouseup"）
+- L5 评价：按某个标准判断现有结果好不好并说明理由，或比较不同做法再做选择（例："现在按钮太多，小学生看不懂，改成只留三个主要按钮"）
+- L6 创造：在题目要求之外提出自己的设计，如原创的主题、玩法、交互或整体风格，并说明它们怎么配合（例："改成星空主题：转盘是星座，抽中稀有奖励时流星划过并解锁对应星座的图鉴"）
+
+判断要点：
+- 只看学生自己想出来的部分。和题目要求重合的内容不算创造，要求写得长不等于层级高。
+- 结合上一轮看：重复上一轮的提问，层级不高于上一轮。
 
 只返回一个数字（1-6），不要解释。`
 
 const MIN_PROMPT_CHARS = 4
+const BRIEF_CHARS = 800
+const PROMPT_CHARS = 1500
 
 const sleep = (ms: number) => new Promise((r) => setTimeout(r, ms))
 
@@ -40,9 +48,45 @@ async function callLlm(content: string): Promise<number | null> {
   }
 }
 
+// 题目要求去掉图片、压成一段，太长的截掉
+function briefOf(content: string) {
+  const text = content
+    .replace(/!\[[^\]]*\]\([^)]*\)/g, "")
+    .replace(/\s+/g, " ")
+    .trim()
+  return text.length > BRIEF_CHARS ? text.slice(0, BRIEF_CHARS) + "……" : text
+}
+
+function clip(text: string) {
+  return text.length > PROMPT_CHARS ? text.slice(0, PROMPT_CHARS) + "……" : text
+}
+
+// 发给模型的：题目要求、同一对话里上一轮的提问（引导对话只和引导对话比）、这一轮的提问
+function classifyInput(msg: { id: number; content: string; source: string; conversation_id: string; created: string }) {
+  const task = get<{ title: string; content: string }>(
+    `SELECT t.title, t.content FROM prompt_conversation c JOIN task_task t ON t.id = c.task_id WHERE c.id = $conv`,
+    { conv: msg.conversation_id },
+  )
+  const prev = get<{ content: string }>(
+    `SELECT content FROM prompt_message
+     WHERE conversation_id = $conv AND role = 'user' AND source != 'edit'
+       AND (source = 'guidance') = $guidance
+       AND (created < $created OR (created = $created AND id < $id))
+     ORDER BY created DESC, id DESC LIMIT 1`,
+    { conv: msg.conversation_id, guidance: msg.source === "guidance", created: msg.created, id: msg.id },
+  )
+  return [
+    `题目：${task?.title ?? "（未知）"}`,
+    `题目要求：${task ? briefOf(task.content) : "（无）"}`,
+    `上一轮提问：${prev ? clip(prev.content) : "（这是第一轮）"}`,
+    `这一轮提问：${clip(msg.content)}`,
+  ].join("\n\n")
+}
+
 export async function classifyMessage(messageId: number): Promise<number | null> {
-  const msg = get<{ content: string }>(
-    `SELECT content FROM prompt_message WHERE id = $id AND role = 'user' AND source != 'edit'`,
+  const msg = get<{ id: number; content: string; source: string; conversation_id: string; created: string }>(
+    `SELECT id, content, source, conversation_id, created FROM prompt_message
+     WHERE id = $id AND role = 'user' AND source != 'edit'`,
     { id: messageId },
   )
   if (!msg) return null
@@ -51,7 +95,7 @@ export async function classifyMessage(messageId: number): Promise<number | null>
     run(`UPDATE prompt_message SET prompt_level = NULL WHERE id = $id`, { id: messageId })
     return null
   }
-  const level = await callLlm(msg.content)
+  const level = await callLlm(classifyInput(msg))
   if (level !== null) {
     run(`UPDATE prompt_message SET prompt_level = $level WHERE id = $id`, {
       level,