k2f-research-deck

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

K2F Research Deck — AI-Native Aesthetic Design

K2F 研究演示文稿 — AI 原生美学设计

Purpose

目的

Teach agents to blueprint research decks before layout: one scientific claim per slide, traceable evidence, explicit caveats before the close, and no numbers or figures redrawn from memory.

教授 Agent 在排版前规划研究演示文稿：每页呈现一个科学主张、确保证据可追溯、在收尾前明确说明注意事项，并且不得凭记忆重绘数字或图表。

Built on K2F

基于 K2F 构建

This skill produces research presentations as

.K2F

packages. Follow the k2f skill for authoring: unpack or init a workspace, edit the semantic tree, pack/verify, and export. K2F keeps agents on meaning (slide roles, evidence, visual intent)—not coordinates—while the engine compiles layout, typography, and theme into consistent, projector-ready output.

Before you build, install the core skill:

npx skills add suzheng/k2f --skill k2f

此 skill 将研究演示文稿生成为

.K2F

包。编写时遵循 k2f skill：解包或初始化工作区、编辑语义树、打包/验证并导出。K2F 让 Agent 专注于内容含义（页面角色、证据、视觉意图），而非坐标；引擎则将布局、排版和主题编译为一致且适合投影的输出。

开始构建前，安装核心 skill：

npx skills add suzheng/k2f --skill k2f

。

When to Use

使用场景

Preparing a lab meeting or PI update from a paper, preprint, or experiment log.
Building a grant or internal review deck that must prove novelty without overclaiming.
Converting dense manuscripts into a 6–15 slide talk with faithful metrics and figures.
Rebuilding generic “paper summary” AI slides into a skimmable evidence story.

根据论文、预印本或实验记录准备实验室会议或 PI 汇报。
制作既要证明创新性又不能过度宣称的基金申请或内部评审演示文稿。
将内容密集的手稿转换为包含准确指标和图表的 6–15 页演讲。
将泛化的“论文摘要”AI 幻灯片重塑为便于快速浏览的证据叙事。

Workflow

工作流

Step 1. Generate the content_and_design.md

第 1 步：生成 content_and_design.md

Single blueprint for semantic content and aesthetic intent. Use labeled blocks per slide (role, headline, body, visual type, asset notes) that the k2f skill can map later.

Sub-steps

Brief frame — Audience, decision target (inform, approve next step, defend novelty), time budget, language. Live talk = sparse bullets; async deck = self-explanatory headlines per slide.
Source-of-truth audit — List abstract, methods sketch, figures/tables, exact results, citations, limitations. Tag
```
NEEDS USER INPUT
```
for gaps; never invent stats, n, or citations.
Evidence spine — Arc before titles: stakes → gap → question/hypothesis → design → results (strongest first) → interpretation → caveats → next step/ask. External reviewers: contributions early; lab updates: lead with freshest result.
Slide architecture (include / cut) — Target 6–15 slides; merge or omit when inputs are thin:
- Title (paper metadata) → Context/problem → Gap/question → Approach overview → Methods (diagram-led) → Main results (split quant vs qual if both exist) → Mechanism/interpretation → Limitations → Next steps → (optional) key references.
- Cut literature surveys, duplicated abstract+introduction, method prose that belongs in backup, and extra ablation slides that do not change the decision.
- Split when two claims share one slide (problem+hypothesis, multiple unrelated figures, or results+interpretation).
Per-slide content contract — For each slide specify:
- Headline: conclusion, not section label (“Knockdown reduces latency 42% at n=12” not “Results”).
- Body: max 3–5 bullets or one visual block; ~2 lines per bullet; no orphan single-word lines.
- Visual type: methods pipeline, results chart, comparison matrix, qualitative grid, equation focus, contribution list, agenda/outline—prefer chart/table/figure intent over paragraphs for data.
- Asset fidelity: source figure/table (figure #, caption) vs schematic only; do not bulletize tables that should display as tables.
Numbers — One unit system; matching metrics across slides; n/CI/test on hero numbers;
```
[VERIFY]
```
if uncertain.
Design intent — Tone: restrained, evidence-led. Hierarchy: title largest; one hero figure/metric/equation per slide. Density: one message per slide; section breaks between movements. Palette: calm academic, one accent, projection-safe contrast.

Rules & constraints

One-claim gate: if the headline needs “and,” split the slide or demote a thread to backup.
Fidelity gate: quantitative slides use user-supplied numbers or explicit TBD—not rounded or “typical” values.
Figure gate: main plots/tables use source assets or explicit chart intent—not prose-redrawn data.
Caveat gate: limitations or confounds before the closing ask.
Novelty gate: one scannable “what’s new”; cut known background.

Common failure modes

Ten slides of background, one of results → reorder spine; promote results and design earlier.
Methods as paragraph walls → one pipeline diagram + 3 bullets max.
Garbled or invented statistics → revert to source table/chart intent or mark TBD.
Missing limitations → add explicit caveats slide; avoids PI “what about confound X?” derail.
Duplicate paper abstract on multiple slides → single context slide, then gap.
Crowded multi-panel figures → one panel per slide or callouts in the blueprint.
Overclaiming headlines → match data; speculation under “open questions.”

Quality checklist

Decision target and audience stated at top of the blueprint.
Every slide has one headline claim and a declared visual type.
Metrics and assets trace to inputs, TBD, or
```
[VERIFY]
```
.
Caveats and next steps before the final slide.
6–15 slides (or justified); design intent covers tone and hierarchy for the deck.

使用单一蓝图定义语义内容和美学意图。每页使用带标签的内容块（角色、标题、正文、视觉类型、素材说明），以便 k2f skill 后续映射。

子步骤

简报框架 — 明确受众、决策目标（提供信息、批准下一步、论证创新性）、时间预算和语言。现场演讲 = 精简要点；异步演示文稿 = 每页标题均可独立说明问题。
事实来源审计 — 列出摘要、方法概述、图表、精确结果、引用和局限性。对于缺失内容标记
```
NEEDS USER INPUT
```
；绝不虚构统计数据、n 或引用。
证据主线 — 先规划叙事弧线，再确定标题：重要性 → 缺口 → 问题/假设 → 设计 → 结果（最强结果优先）→ 解读 → 注意事项 → 下一步/请求。面对外部评审时，尽早呈现贡献；实验室内部汇报时，以最新结果开场。
页面架构（保留/删减） — 目标为 6–15 页；输入内容不足时应合并或省略：
- 标题（论文元数据）→ 背景/问题 → 缺口/问题 → 方法概览 → 方法（以图示为主）→ 主要结果（若同时存在定量和定性结果则分开）→ 机制/解读 → 局限性 → 下一步 →（可选）关键参考文献。
- 删减文献综述、重复的摘要和引言、应放入备用页的方法文字，以及不会影响决策的额外消融实验页面。
- 拆分一页承载两个主张的情况（问题+假设、多个互不相关的图表，或结果+解读）。
单页内容契约 — 为每页指定：
- 标题：结论而非章节名称（例如使用“敲低使延迟在 n=12 时降低 42%”，而不是“结果”）。
- 正文：最多 3–5 个要点或一个视觉内容块；每个要点约 2 行；不要出现孤立的单词行。
- 视觉类型：方法流程图、结果图表、对比矩阵、定性分析网格、公式聚焦、贡献列表、议程/大纲；对于数据，优先采用图表/表格/插图意图，而不是段落。
- 素材保真度：使用源图表（图号、图注）还是仅使用示意图；不要将本应以表格形式展示的表格改写成要点。
数字 — 统一单位制；各页指标保持一致；核心数字应标注 n/CI/检验结果；不确定时使用
```
[VERIFY]
```
。
设计意图 — 基调：克制、以证据为导向。层级：标题最大；每页仅设置一个核心图表、指标或公式。密度：每页传达一个信息；在不同叙事阶段之间设置章节分隔。配色：沉静的学术风格，使用单一强调色，并确保投影时的对比度。

规则与约束

单一主张门槛：如果标题需要使用“和”，应拆分页面，或将该线索降级至备用页。
保真度门槛：定量页面只能使用用户提供的数字或明确标记的 TBD，不得使用四舍五入或“典型”数值。
图表门槛：主要图表和表格应使用源素材或明确的图表意图，不得用文字重绘数据。
注意事项门槛：在最终请求前呈现局限性和混杂因素。
创新性门槛：清晰呈现一项可快速识别的“新内容”；删减已知背景。

常见失败模式

10 页背景、1 页结果 → 重新排序主线；提前突出结果和设计。
方法部分变成大段文字 → 使用一张流程图和最多 3 个要点。
统计数据混乱或被虚构 → 恢复为源表格/图表意图，或标记为 TBD。
缺少局限性 → 添加明确的注意事项页面，避免 PI 因“混杂因素 X 怎么办？”等问题而偏离重点。
多页重复论文摘要 → 保留一个背景页面，然后呈现缺口。
多面板图表过于拥挤 → 每页展示一个面板，或在蓝图中标注重点区域。
标题过度宣称 → 与数据保持一致；推测内容应归入“开放问题”。

质量检查清单

蓝图顶部已说明决策目标和受众。
每页都有一个标题主张和明确的视觉类型。
指标和素材可追溯至输入内容、TBD 或
```
[VERIFY]
```
。
最终页之前已呈现注意事项和下一步计划。
页面数量为 6–15 页（或有合理说明）；设计意图涵盖演示文稿的基调和层级。

Step 2. Use the content_and_design.md and k2f skill to generate a K2F file and export to an appropriate file format.

第 2 步：使用 content_and_design.md 和 k2f skill 生成 K2F 文件并导出为适当的文件格式

Follow k2f to instantiate slides from the blueprint, apply roles from design intent, verify, and export the presentation-ready deliverable.

遵循 k2f 的流程，根据蓝图实例化页面、应用设计意图中的角色、进行验证，并导出可直接用于演示的交付物。