1) 目标与约束

设计原则（强制）

• P1 LLM 输出必须是 结构化 JSON，通过 schema 校验后才能进入下一阶段
• P2 布局由 确定性排版引擎 完成，支持文字测量与溢出回退
• P3 图像只做“背景/插画/装饰/图标”，不承载正文信息
• P4 Repair loop：规则优先，LLM 仅负责“改写/拆页/重组要点”

2) 总体架构与数据流

Inputs (sources + user prompt + optional brandbook/sample deck)
        |
        v
[1] Planner (LLM) ---------> Slide Outline IR (v0)
        |
        v
[2] Style Builder (LLM+rules) -> Theme Tokens
        |
        v
[3] Composer (rules+LLM opt) -> Component Tree IR (v1)
        |
        v
[4] Layout Engine (deterministic) -> Layout IR (bboxes + line breaks)
        |
        v
[5] Renderer (HTML) -> deck.html
        |
        v
[6] Validators (rules) -> issues
        |
        v
Repair Loop:
  - rule fixes (fast) OR
  - LLM rewrite/split/template swap
        |
        v
Re-layout -> Re-validate -> Done / Fallback

3) 模型分工与路由策略

3.1 推理模型职责

3.2 图像模型职责

3.3 路由策略（可落地的最小规则）

// Router 伪代码（思路）
task = {type:"planner"|"rewrite"|"extract"|"arbiter", difficulty, latency_budget}
if task.type=="planner" and task.difficulty=="high": use GPT_5_2
else if task.type=="rewrite": use Claude_Opus
else if task.type=="extract": use Gemini_3
else: use GPT_5_2

// 可选：并行 + 仲裁
candidates = run_parallel([primary, secondary])
best = pick_by(schema_valid && fewer_issues && shorter_text && preserves_keys)

4) 统一输出契约（Schema/JSON）

4.1 Slide IR（v1）最小字段

{
  "deck": {
    "meta": {
      "title": "string",
      "language": "zh-CN",
      "mode": "presenter|detailed",
      "aspect": "16:9",
      "pageSize": { "w": 1920, "h": 1080 }
    },
    "themeRef": "theme.default",
    "slides": [
      {
        "id": "s1",
        "intent": "cover|section|concept|comparison|process|timeline|data|quote|summary",
        "title": "string",
        "subtitle": "string?",
        "bullets": ["string"],
        "callouts": ["string?"],
        "notes": "string?",
        "assets": [
          { "role":"bg|illustration|icon|chart", "ref":"assetId?", "styleHint":"string?" }
        ],
        "constraints": { "maxBullets": 5, "maxWordsPerBullet": 14 }
      }
    ]
  }
}

4.2 Layout IR（输出给 Renderer）

{
  "page": { "w":1920, "h":1080, "safe":48, "grid":8 },
  "slides": [
    {
      "id":"s1",
      "nodes":[
        {
          "nodeId":"title",
          "type":"text|list|image|shape|chart",
          "bbox":{ "x":96, "y":96, "w":1200, "h":140 },
          "styleRole":"h1|h2|body|caption",
          "text":"...",
          "font":{ "family":"Inter", "size":44, "weight":700, "lineHeight":1.15 },
          "lines":["..."]  // 可选：已换行结果
        }
      ]
    }
  ]
}

4.3 Issues（Validators 输出）

{
  "issues":[
    {"type":"overflow","slideId":"s2","nodeId":"bullets","excessPx":72,"severity":"high"},
    {"type":"overlap","slideId":"s3","nodeA":"title","nodeB":"visual","severity":"high"},
    {"type":"contrast","slideId":"s1","nodeId":"title","ratio":2.3,"severity":"med"}
  ]
}

5) Prompt 规范与模板

5.1 Planner Prompt（给 GPT-5.2 / Gemini）

{
  "task":"planner",
  "input":{
    "mode":"presenter|detailed",
    "language":"zh-CN",
    "audience":"string",
    "goal":"string",
    "sources":[{"id":"p1","text":"..."}],
    "constraints":{
      "slidesMin":8,
      "slidesMax":12,
      "maxBulletsPresenter":4,
      "maxBulletsDetailed":6,
      "maxWordsPerBulletPresenter":12,
      "maxWordsPerBulletDetailed":18
    }
  },
  "output_format":"ONLY_JSON_MATCHING_SCHEMA(deck_ir_v1)"
}

5.2 Rewrite Prompt（给 Claude）

{
  "task":"rewrite",
  "input":{
    "target":"shorter|clearer|more_formal|more_playful",
    "text":"原 bullet/段落",
    "must_keep":["关键词A","数字B","结论C"],
    "limits":{"maxWords":12}
  },
  "output_format":"ONLY_JSON {\"text\":\"...\"}"
}

5.3 Repair Prompt（给 GPT-5.2 仲裁）

{
  "task":"repair",
  "input":{
    "deck_ir":"...",
    "issues":"...",
    "allowed_actions":[
      "rewrite(nodeId,target)",
      "trim_bullets(nodeId,keepTopK)",
      "swap_template(slideId,templateId)",
      "split_slide(slideId,intoN)"
    ],
    "hard_limits":{
      "minFontBody":14,
      "maxIterations":4
    }
  },
  "output_format":"ONLY_JSON {\"actions\":[...]}"

6) 风格系统（Tokens）

6.1 Tokens 必备字段

{
  "themeId":"theme.brandA",
  "palette":{"bg":"#0B1220","text":"#E6EDF3","primary":"#7AA2FF","surface":"#101826","line":"#22324A"},
  "typography":{"titleFont":"Inter","bodyFont":"Inter","scale":{"h1":44,"h2":32,"body":18,"small":14}},
  "spacing":{"grid":8,"safeMargin":48,"gap":16,"lineHeight":1.25},
  "shape":{"radius":16,"stroke":1,"shadow":"soft"},
  "imagery":{"style":"flat|handdrawn|realistic|chalk|manga","bgPattern":"none|grain|dots"},
  "charts":{"axis":"minimal|full","labels":"compact|full"}
}

6.2 “任意风格”的工程做法

• 先做 预置主题库（10~30 套 tokens），把常见风格映射到 tokens（极简/科技/复古/手绘/黑板等）
• 再做 brandbook 抽取（可用推理模型解析文本规则）→ tokens 覆盖
• 风格只改变 tokens，不改变模板与布局逻辑（保证稳定）

7) 模板库与组件树

7.1 Template DSL（结构+约束）

{
  "templateId":"concept.2col.imageRight",
  "intent":"concept",
  "grid":{"cols":12,"rows":12,"safe":48},
  "slots":[
    {"name":"title","type":"text","area":[1,1,7,2],"styleRole":"h2"},
    {"name":"bullets","type":"list","area":[1,3,7,8],"styleRole":"body","maxLines":10},
    {"name":"visual","type":"media","area":[8,2,11,10],"fit":"cover","minVisible":0.6}
  ],
  "constraints":{"align":"baselineGrid","minGap":16,"noOverlap":true}
}

7.2 组件树组装规则

• 按 intent 先选模板候选（3~5 个），再根据内容密度与媒体可用性打分
• slot 填充时标记优先级：标题（最高）> 关键结论 > bullets > 装饰
• 为 Repair 留后门：bullets 支持 trim，visual 支持缩放/替换，必要时 split

8) 确定性排版引擎（关键）

8.1 必做能力

• 文字测量：给定字体/字号/宽度，计算换行与高度
• 基线网格：元素吸附 grid（8px），保证齐整
• 溢出回退：rewrite → downscale → trim → swap template → split
• 下限约束：正文字号不低于 minFontBody（例如 14）

8.2 回退策略（建议顺序）

1) Rewrite（Claude 或 GPT-5.2）：更短更清晰，不丢 must_keep
2) Downscale（规则）：降字号到下限
3) Trim（规则）：bullets 保留 top-k，其余拆页
4) Swap template（规则）：换更“文字友好”模板
5) Split slide（GPT-5.2 仲裁 + 规则执行）：拆成两页或更多
6) Fallback：summary/section 兜底模板

9) 图像生成管线（NanoBanana Pro / Image 1.5）

9.1 使用边界

• 允许：背景纹理、插画、图标、抽象装饰、无字素材图
• 不建议：把标题/正文写进图片（除非你们明确接受不可编辑）
• 推荐构图：为叠字预留留白区；背景不过度花（避免对比度问题）

9.2 图像提示词输入（统一格式）

{
  "task":"image_generate",
  "model":"nanobanana_pro|image_1_5",
  "input":{
    "slideId":"s3",
    "intent":"concept",
    "imagery":{
      "style":"flat|handdrawn|realistic|chalk|manga",
      "mood":"calm|bold|playful|serious",
      "composition":"leave whitespace on left for text",
      "avoid":"no text, no watermark, no logos"
    },
    "palette":{"primary":"#7AA2FF","bg":"#0B1220","accent":"#34D399"},
    "size":{"w":1920,"h":1080},
    "transparent": false
  },
  "output_format":"image_binary_or_url"
}

9.3 选型建议（工程规则）

• 需要强风格化一致性（一整套 deck 插画风一致）：优先 NanoBanana Pro
• 需要通用写实素材/背景变体：优先 Image 1.5
• 图像失败/不合规：降级为纯色/渐变/几何纹理（规则生成，无需模型）

10) Validators + Repair Loop

10.1 Validators（规则）

10.2 Repair Loop（收敛机制）

• 最多迭代 3~5 次；每次都必须重新 Layout + Validate
• 规则修复优先；LLM 只做文本改写/拆页建议/模板仲裁
• 若不收敛：强制 split 或 fallback（summary）保证输出稳定

{
  "actions":[
    {"type":"trim_bullets","slideId":"s2","nodeId":"bullets","keepTopK":4},
    {"type":"llm_rewrite","model":"claude_opus_4_5","slideId":"s2","nodeId":"bullets","target":"shorter"},
    {"type":"swap_template","slideId":"s3","templateId":"concept.singleCol.textHeavy"}
  ]
}

11) HTML 渲染与导出

11.1 HTML 渲染建议

• 每页一个 <section class="slide">，固定尺寸（推荐 1920x1080）
• tokens → CSS variables；styleRole → class；bbox → inline style（left/top/width/height）
• 导出友好：确保字体加载完成再截图/打印

<section class="slide" style="width:1920px;height:1080px">
  <div class="node title h1" style="left:96px;top:96px;width:1200px;height:140px">...</div>
  <ul class="node bullets body" style="left:96px;top:260px;width:920px;height:700px">...</ul>
  <img class="node visual" style="left:1100px;top:200px;width:720px;height:760px;object-fit:cover" />
</section>

11.2 导出 PDF（推荐）

• Playwright/Puppeteer：加载 HTML → document.fonts.ready → pdf/逐页截图
• 打印背景：printBackground:true

12) 工程化：服务拆分、缓存、日志、评测

12.1 服务拆分（建议）

/api/plan        (LLM: GPT-5.2 or Gemini)  -> deck_ir_v1
/api/style       (rules + LLM optional)    -> tokens
/api/compose     (rules + LLM optional)    -> deck_ir_v1 (slots filled)
/api/layout      (deterministic)           -> layout_ir
/api/validate    (rules)                   -> issues
/api/repair      (LLM arbiter + rules)     -> actions
/api/image       (NanoBanana/Image1.5)     -> asset
/api/render/html (deterministic)           -> html

12.2 缓存键（强烈建议）

• LLM：hash(prompt + schemaVersion + model + temperature)
• 文字测量：hash(fontFamily + size + width + text)
• 图片：hash(tokens + intent + prompt + size + model)

12.3 评测与回归（必须做）

• golden cases：长标题、超多 bullets、无图、全图、CJK、RTL
• 属性测试：随机文本长度/密度，断言“不溢出/不重叠”
• 视觉回归：截图 diff（阈值）

13) MVP 交付清单（按最短路径）

附录：示例 Schema/接口

A) Model Router 请求体（统一）

{
  "model":"gpt_5_2_thinking|claude_opus_4_5|gemini_3_pro|nanobanana_pro|image_1_5",
  "task":"planner|rewrite|repair|extract|image_generate",
  "schema":"deck_ir_v1|repair_actions_v1|rewrite_v1|...",
  "input":{ "...": "..." },
  "settings":{
    "temperature": 0.2,
    "maxTokens": 4000,
    "timeoutMs": 45000
  }
}

B) Repair Actions Schema（简化）

{
  "actions":[
    {"type":"llm_rewrite","model":"claude_opus_4_5","slideId":"s2","nodeId":"bullets","target":"shorter"},
    {"type":"trim_bullets","slideId":"s2","nodeId":"bullets","keepTopK":4},
    {"type":"swap_template","slideId":"s3","templateId":"process.vertical.clean"},
    {"type":"split_slide","slideId":"s4","intoN":2}
  ]
}