「质胜文则野，文胜质则史。文质彬彬，然后君子。」——《论语 · 雍也》
品牌释义：质 = 内容、论证、策略（Strategy → IR 管线）；文 = 渲染、视觉、呈现（Renderer）。AI 演示的常见失败是"文胜质"——形式压过内容。「文质」先把"质"立住，再让"文"恰如其分。
不是又一个 PPT 模板生成器——而是一套把演示科学（认知科学 / 传播学 / 演示研究）与 Web 演示工程打通的 AI Skill。给定复杂内容、真实受众、沟通目标与场景，它产出的不是"幻灯片文件"，而是一个可现场运行、可分支深潜、可动态调整的 Presentation Application（演示应用）。
Not another slide-template generator — an AI Skill that fuses presentation science (cognitive science, communication, presentation research) with web presentation engineering. Given complex content, a real audience, a communication goal and a situation, it produces not "slide files" but a runnable, branchable, adaptive Presentation Application.
官网 / Website：https://wenzhi.mizzlelover.xyz · 在线 Demo / Live Demo：https://wenzhi.mizzlelover.xyz/examples/demo/（本地运行：`examples/demo/index.html`）
| 常见工具 | 本系统 |
|---|---|
| 输入 → 直接生成 HTML 页面 | 输入 → 演示策略 → 论证架构 → Scene 规划 → Presentation IR → 渲染 → QA（禁止一步到位） |
| 通用规则："每页 6 行""标题必须问句" | 证据分级（A–E）+ 边界条件：规则必须带机制、语境、证据、例外 |
| 线性翻页器 | Presentation Graph：核心路径 / 可选路径 / Deep Dive / 证据附录 / 现场分支 |
| 只有"放映"一种形态 | One Content Model, 三种 Mode：Stage（讲演）/ Reader（自读）/ Print（打印） |
| 动画=装饰 | 语义动效（REVEAL/FOCUS/CAUSE…），讲者可 pause/seek/skip |
| 依赖 CDN、现场断网即崩 | 离线优先：整目录打包，Stable > Fancy |
# 1. 安装到你的 AI 编程环境（Claude Code / Codex / OpenCode / Kimi Code）
cp -r wenzhi ~/.config/agents/skills/ # 用户级
# 或项目级：cp -r wenzhi your-project/.agents/skills/
# 2. 看一眼可运行的完整示例（零依赖，浏览器直接打开）
open examples/demo/index.html
# 快捷键：→/Space 推进 · ← 回退 · O 总览 · M 静帧 · / 搜索 · P 讲者台 · R 阅读模式 · ? 帮助
# 3. 校验一份 Presentation IR
python3 scripts/validate_ir.py examples/demo/presentation.ir.json各 harness 的详细接入方式见 adapters/ · See adapters/ for Claude Code / Codex / OpenCode setup.
产物自带完整的演示操作层：总览平铺、静帧开关、讲者控制台、自读模式、搜索与帮助——所有键位在演示界面右下角常驻隐晦提示，? 随时唤起完整面板。
源材料 Source Material
↓ analyze_source_content 内容理解 + 证据标注
演示策略 Presentation Strategy 情境判定(20类) → 受众建模 → 目标动词
↓ build_argument_map 论证树 + 六查（无证据断言/逻辑跳跃/过度声称…）
Scene 规划 Scene Plan 每幕一个 Cognitive Job · 认知负荷曲线 · 非线性 Graph
↓ build_presentation_ir 唯一中间表示（validate_ir.py 强制校验）
HTML 渲染 Rendering 语义组件 + design tokens + L0–L6 最低充分渲染
↓ review_presentation 18 维评测 · 视觉/Runtime QA · 浏览器实测
演示应用 Presentation App 离线包：Stage + Reader + Print 三模式
- 35+ 来源 Seed Corpus：Mayer（多媒体学习）、Sweller（认知负荷）、Alley（Assertion-Evidence）、Cleveland & McGill（图形感知）、Petty & Cacioppo（ELM）、Minto、Tufte、Duarte……全部登记证据等级与边界条件（EVIDENCE.md）。
- 知识图谱节点：每个方法论点含机制 / 支持证据 / 冲突证据 / 边界 / 失效条件 / 常见误读（knowledge/）。
- 经验假设层：用户经验不直接成为规则，先登记为 Practitioner Hypothesis 再验证（registry）。
- 反模式库：通用规则谬误 / 装饰优先 / 技术滥用（anti_patterns）。
- 对抗测试：当用户说"字越少越好""每页加炫酷动画""所有数据做 3D"时，系统知道如何有理有据地回应（adversarial.md）。
SKILL.md — Skill 路由（progressive disclosure）
knowledge/ — 演示知识图谱 + 证据分级 + 经验假设层 + 反模式
schemas/ — IR / Scene / 知识节点 / 来源 YAML Schema
workflows/ — 9 个标准工作流
runtime/ — 零依赖 HTML Presentation Runtime（实测通过）
components/ — 语义组件（信息结构，非模板）
themes/ — design tokens 主题 ×5
visualization/ — ECharts/D3/SVG/Mermaid 选型指南
evals/ — 18 维 rubric · 100 测试案例 · 10 对抗测试
scripts/ — validate_ir.py（IR 校验器）
examples/demo/ — 完整可运行示例（自我推介，在线运行：/examples/demo/）
adapters/ — Claude Code / Codex / OpenCode 接入
index.html — 官方宣传页（GitHub Pages 站点首页）
assets/ — 站点资源（公众号官方物料等）
详细设计：ARCHITECTURE.md · METHODOLOGY.md · RUNTIME.md · DESIGN_SYSTEM.md · EVALS.md · FINAL_REPORT.md
Wenzhi · Web-native Presentation Intelligence is an agent skill that builds HTML presentation applications — not slide files. It fuses presentation science (cognitive science, communication research, presentation studies) with web presentation engineering: given complex content, a real audience, a communication goal and a situation, it produces a runnable, branchable, adaptive presentation app with Stage / Reader / Print modes.
The name 文质 (wénzhì) comes from the Analects of Confucius: "When substance exceeds refinement, one becomes crude; when refinement exceeds substance, one becomes a pedant. Only when substance and refinement are balanced does one become a person of quality." — 质 (substance) is content, argument and strategy; 文 (refinement) is rendering, visuals and delivery. AI presentations usually fail as "refinement over substance"; Wenzhi secures substance first, then lets refinement serve it.
| Common tools | Wenzhi |
|---|---|
| Input → HTML in one shot | Input → strategy → argument map → scene plan → Presentation IR → render → QA (one-shot generation is forbidden) |
| Universal rules ("max 6 lines per slide") | Evidence-graded rules (A–E) with boundary conditions: every rule carries mechanism, context, evidence and exceptions |
| Linear page-flipper | Presentation Graph: core path / optional paths / deep dives / evidence appendix / live branching |
| Show mode only | One content model, three modes: Stage (live talk) / Reader (self-reading) / Print |
| Animation as decoration | Semantic motion (REVEAL / FOCUS / CAUSE …), presenter can pause / seek / skip |
| CDN-dependent, dies offline | Offline-first: the whole folder runs anywhere; stability over fancy |
# 1. Install into your AI coding environment (Claude Code / Codex / OpenCode / Kimi Code)
cp -r wenzhi ~/.config/agents/skills/ # user-level
# or project-level: cp -r wenzhi your-project/.agents/skills/
# 2. Run the complete demo (zero dependencies, opens in any browser)
open examples/demo/index.html
# Keys: →/Space advance · ← back · O overview · M still-frame · / search
# P presenter console · R reader mode · ? help · Esc dive-return
# 3. Validate a Presentation IR
python3 scripts/validate_ir.py examples/demo/presentation.ir.jsonEvery generated deck ships with a full presentation layer — overview grid, still-frame switch, presenter console, reader mode, search and help — with an unobtrusive always-on hint in the corner and a complete ? panel. Generate a matching key-card for your own deck with scripts/gen_keys_card.py.
Source Material
↓ analyze_source_content content understanding + evidence tagging
Presentation Strategy situation (20 types) → audience model → goal verb
↓ build_argument_map argument tree + six checks (unsupported claim, logical gap, overclaim…)
Scene Plan one cognitive job per scene · load curve · non-linear graph
↓ build_presentation_ir single source of truth (enforced by validate_ir.py)
Rendering semantic components + design tokens + L0–L6 minimum-sufficient rendering
↓ review_presentation 18-dimension rubric · visual/runtime QA · real-browser testing
Presentation App offline bundle: Stage + Reader + Print
- 35+ graded sources: Mayer (multimedia learning), Sweller (cognitive load), Alley (assertion–evidence), Cleveland & McGill (graphical perception), Petty & Cacioppo (ELM), Minto, Tufte, Duarte — each registered with evidence grade and boundary conditions (EVIDENCE.md).
- Knowledge-graph nodes: every method carries mechanism / supporting evidence / conflicting evidence / boundaries / failure conditions / common misreadings (knowledge/).
- Practitioner-hypothesis layer: field experience is quarantined as a hypothesis until validated — it never becomes a rule directly (registry).
- Anti-pattern library: universal-rule fallacies, decoration-first, technology abuse (anti_patterns).
- Adversarial benchmarks: when asked for "cool animation on every page" or "3D everything", the system answers with mechanism-level reasoning (adversarial.md).
SKILL.md — skill router (progressive disclosure)
knowledge/ — presentation knowledge graph + evidence grading + hypotheses + anti-patterns
schemas/ — YAML schemas: IR / scene / knowledge node / source
workflows/ — 9 standard workflows
runtime/ — zero-dependency HTML presentation runtime (browser-tested)
components/ — semantic components (information structures, not templates)
themes/ — design-token themes ×5
visualization/ — ECharts / D3 / SVG / Mermaid selection guides
evals/ — 18-dimension rubric · 100 test cases · 10 adversarial tests
scripts/ — validate_ir.py (IR validator) · gen_keys_card.py (keyboard-map card)
examples/demo/ — complete runnable demo (self-introduction deck, live at /examples/demo/)
adapters/ — Claude Code / Codex / OpenCode integration
index.html — official promo page (GitHub Pages site root)
assets/ — site assets (WeChat official material, key cards)
Detailed design: ARCHITECTURE.md · METHODOLOGY.md · RUNTIME.md · DESIGN_SYSTEM.md · EVALS.md · FINAL_REPORT.md
谁是专家 — 关注 AI × 演示 × 知识工程的实践者。
- 小红书：谁是专家
- X (Twitter)：@dboy_yi2025
- GitHub：mizzlelover
- 微信公众号「谁是专家」——扫码关注，第一时间获取更新：
MIT © 2026 谁是专家