AI Presentation Prompt Tips (2026): Make Slides Feel Less AI-Generated

If you throw raw material directly into an AI presentation tool, nine times out of ten you get the canned AI deck: blue-purple gradients, full-slide bullet lists, and generic glowing visuals. A three-step workflow fixes this: discuss the outline with Claude or ChatGPT → generate with Gamma / NotebookLM → manually polish colors and images. This article collects prompt patterns you can copy and adapt.

Why AI Presentations Look Obviously AI-Made

Open any AI-generated deck and it usually looks like this: dark background with neon gradients, six or seven bullet points per slide, and glowing abstract geometry everywhere. The problem is the prompt.

AI visual models were trained on lots of “techy” Dribbble-style work, so prompts like “modern” or “futuristic” fall straight into that visual trap. Language models also naturally compress information, so they squeeze every key point onto one slide. The result is a beautiful Word document, not a presentation.

To escape this, instructions need to be more precise.

Three-Step Workflow: Discuss, Generate, Polish

Core idea: let the AI that is good at thinking do the thinking, and the AI that is good at layout do the layout.

Step 1: Discuss the outline with a strong model

Do not let Gamma or NotebookLM guess the important points of a 30-page report by itself. Start a Claude or ChatGPT conversation, paste the source material, and use this prompt:

你是資深管理顧問。分析以下資料，萃取最關鍵的策略洞察，
整理成 10-12 張投影片的大綱。每張投影片包含：
1. 一個結論式標題（例如「Q3 營收超標 12%，三個關鍵推手」而非「Q3 營收分析」）
2. 三個支持論點（bullet point，各不超過 15 字）
3. 建議的視覺元素（圖表類型或照片方向）
不要寫完整段落。不要加入原始資料沒有的內容。

This step lets the stronger reasoning model filter information and arrange the logic. Spending 10 minutes discussing here improves generation quality noticeably.

Step 2: Feed the outline into Gamma or NotebookLM

After you have the outline, paste it directly into the presentation tool. For Gamma, paste it into the prompt field. For NotebookLM, save the outline as a text file, upload it as a source, then choose Slide Deck in the Studio panel.

The key is to include style instructions during generation, described below, instead of letting the tool use its default style.

Step 3: Manual polish

The AI first draft usually reaches 70-80%. The rest is manual:

• Colors: replace AI default gradients with brand colors or low-saturation professional palettes
• Fonts: use audience-appropriate fonts such as Noto Sans for CJK decks; avoid random system defaults
• Images: replace AI abstract art with your own photos or regenerate images in Gemini
• Cut text: if a slide has more than four bullets, cut; if a bullet has more than 20 words, rewrite

Gamma Prompt Examples: Copy and Adapt

Gamma has no hidden syntax. It is natural language plus explicit constraints.

Example 1: Product Introduction Deck for Clients

請為 [產品名稱] 製作一份 10 頁的產品介紹簡報。
目標受眾：時間有限的企業 IT 主管。
結構：問題 → 解法 → 3 個關鍵特色 → 含量化 ROI 的案例研究 → 定價 → 下一步行動。
限制：每頁最多 4 個 bullet，每個 bullet 不超過 12 個字。
設計：乾淨極簡風，配色使用 #1A365D + #F7FAFC + 一個強調色 #38B2AC。使用 sans-serif 字體，保留大量留白，採不對稱版面。不要使用漸層、不要霓虹色、不要放握手的人物素材照。
重點：所有文字都必須是自然流暢的繁體中文。

Example 2: Internal Weekly Report for Managers

請做一份 7 頁的每週狀態報告簡報。
章節：Executive Summary（1 頁）→ 含 YoY 比較圖的關鍵指標 → 本週亮點 → 卡點與風險 → 下週優先事項 → 所需資源 → Appendix。
語氣：以數據為主、精簡、不用 buzzword。每頁只傳達一個核心訊息，最多 3 個 bullet。
配色：navy #001F3F + 白色 + 強調色 teal #00BFFF。只用純色塊，不要漸層，不要裝飾元素。
所有文字都要是繁體中文。

Example 3: Teaching Deck for Students

請為完全沒接觸過 [主題] 的大學生製作一份 12 頁的教學簡報。
結構：引發好奇的問題 → 核心概念 1-4（每頁一個） → 真實世界案例 → 常見錯誤 → 練習題 → 總結 → 延伸閱讀。
風格：大膽、有趣，但不要幼稚。使用大字體，多用圖示和圖解，少用大段文字。每頁最多 15 個字。
背景：暖白色 #FFF8F0，文字：深灰色 #2D3748。僅用一個強調色 #E53E3E。
適合的地方請加入 step-by-step 視覺說明。
所有文字都要是繁體中文。

Example 4: Investor Pitch Deck

請為 [一句話描述產品] 製作一份 10 頁的 seed-stage pitch deck。
頁面：問題 → 解法 → 市場規模（TAM/SAM/SOM，附圖表） → 成長表現 → 商業模式 → Go-to-Market → 競品格局 → 團隊 → 募資需求與資金用途 → Executive Summary。
語氣：有自信、以數據為主、不講空話。每頁最多 6 個 bullet，並包含一張成長表現圖表。
視覺：專業、寫實，避免 AI 生成的人像。
設計：極簡企業風，navy + 白色 + 一個強調色。
所有文字都要是繁體中文。

Extra tip: after generation, use Gamma’s AI Agent (Cmd+E) for global edits, such as “Make this more concise” or “Apply dark theme to entire deck.”

More Copyable Prompt Examples

These templates are source examples. For your own use, swap the final output-language line for your target language.

Proposal Deck for Clients or Managers

Scenario: a deck meant to persuade someone to approve budget or say yes. The focus is evidence and ROI, not visual prettiness. Best tools: Gamma for direct generation, ChatGPT for outline before manual deck building.

做一份 10 頁的提案簡報，目標是說服 [客戶或主管] 同意 [要推動的事]。
受眾：[職位、決策風格]
結構：
  1. 一句話講清楚要解決的問題（配一個客戶場景）
  2. 現況的痛點（三個，各附一個量化數字）
  3. 解法（用一張圖說明流程）
  4. 三個關鍵優勢（每個一句話，不要形容詞）
  5. 一個同業案例（含前後數據對比）
  6. 時程表（分三個階段）
  7. 投入預算和預期 ROI
  8. 風險和對應做法
  9. 團隊背景（三位關鍵成員）
  10. 下一步行動（兩個選項讓對方挑）
每頁最多 4 個 bullet，每個 bullet 最多 12 字。
配色用深藍 #0F2A44 加米白 #FAF8F3，重點色用橘 #F06A3A。
禁用漸層、禁用霓虹、禁用握手照和西裝人像的素材圖。
字體用思源黑體或 Noto Sans TC，留白留多一點。
所有文字都要是繁體中文。

Tip: put “one peer case study” into the structure first, then fill in real data yourself later. AI will reserve the slide area without inventing numbers.

Teaching Deck for Students or New Hires

Best tools: Gamma, NotebookLM with course material as sources.

用上傳的教材做一份 12 頁的教學簡報，受眾是完全沒碰過 [主題] 的大學生或新人。
結構：
  1. 一個引起好奇的問題或生活場景
  2. 今天要回答的三個問題
  3-7. 核心概念 A、B、C 各拆解（含類比圖、生活例子、常見誤解）
  8. 三個概念怎麼串起來（一張關係圖）
  9. 常見錯誤和怎麼避免
  10. 一個 5 分鐘內能做完的練習
  11. 今天學到什麼（學生自己填的 3 題小測驗）
  12. 延伸閱讀清單
風格：活潑但不幼稚，大字體，多圖少字，每頁不超過 15 個字的正文。
背景用暖白 #FFF8F0，主文字深灰 #2D3748，重點色用橘紅 #E53E3E。
圖片請用彩色鉛筆手繪風，禁止漸層和 3D 渲染。
所有文字都要是繁體中文。

Weekly or Monthly Report for Managers or Teams

做一份 7 頁的週報簡報，受眾是主管，他每週只會花 5 分鐘看。
資料：
  本週完成：[條列三到五項]
  進行中：[條列兩到三項]
  卡住的地方：[條列一到兩項，含原因]
  下週重點：[條列三項]
  需要主管協助：[一項]
結構：
  1. 封面（本週、姓名、日期）
  2. 一句話總結本週
  3. 本週完成項目（配進度條或打勾清單）
  4. 關鍵指標表（含上週數字、本週數字、差異百分比）
  5. 卡住的議題（問題、原因、目前對策）
  6. 下週計畫
  7. 需要協助的一件事（大字放中間）
語氣：直接、數據為主、不要形容詞。
配色用深藍 #001F3F 配白底，重點色用亮藍 #00BFFF。
只用純色，不要漸層，不要裝飾線條。
所有文字都要是繁體中文。

Tip: make “one thing I need help with” its own final slide. Managers will remember it.

Technical Sharing Deck for Engineering Teams

做一份 15 頁的技術分享簡報，主題是 [技術名稱]，受眾是熟悉程式但沒用過這個技術的工程師。
結構：
  1. 標題和分享者資訊
  2. 為什麼會踩進這個技術（一個具體的痛點）
  3. 這個技術在解什麼問題（一張架構圖）
  4. 跟現有方案比起來差在哪（比較表）
  5-6. 核心概念一和二，各一頁
  7. 一段程式碼範例（約 15 行，帶註解）
  8. 實際跑起來的畫面或輸出
  9. 在生產環境踩到的三個坑
  10. 效能數字（before / after 圖表）
  11. 什麼情況不該用這個技術
  12. 遷移成本評估
  13. 延伸閱讀和推薦的兩篇文章
  14. QA 頁（留白）
  15. 聯絡方式
風格：技術簡報的語氣，程式碼字型用 Fira Code 或 JetBrains Mono，正文用思源黑體。
配色用深灰 #1A202C 配亮綠 #48BB78 當重點色，不要漸層和裝飾。
程式碼頁不要 AI 合成的假 code，會自己填進去。
所有解說文字都要是繁體中文，程式碼保留英文。

Tip: explicitly say “I will fill the code myself” so AI leaves a placeholder code block.

Product Introduction Deck for Consumers

做一份 8 頁的產品介紹簡報，產品是 [產品名稱]，受眾是一般消費者。
結構：
  1. 產品名稱加一句話 slogan（情感訴求，不要講規格）
  2. 一個具體的生活場景：使用者有什麼煩惱
  3. 怎麼解決這個煩惱（用大圖配三個關鍵字）
  4. 三個最有感的特色（每個配一張產品照）
  5. 真實使用者的一句話心得（三則）
  6. 規格表（尺寸、重量、材質、保固）
  7. 價格和購買管道
  8. 常見問題的三個答案
風格：溫暖、像朋友推薦、不要像電視購物。
配色用暖米 #F5EFE6 配深棕 #5C3A21，重點色用珊瑚橘 #FF7A59。
圖片大、文字少，每頁最多 3 行正文。
禁用科技感、禁用漸層、禁用霓虹色。
所有文字都要是繁體中文。

Tip: if you do not have user quotes yet, mark them as placeholder text such as “quote pending.” AI will reserve three slots.

NotebookLM Tips: Source Quality Decides Everything

NotebookLM’s presentation feature (Studio → Slide Deck) works very differently. It is strictly based on uploaded sources. Source quality matters more than prompt tricks.

How to Organize Sources

1. Upload only cleaned documents, not a raw 200-page report. First use Claude to extract key points, save a clean text file, then upload that.
2. 5-15 sources is the best range. Too many creates noise.
3. Add a brand guide PDF as a visual anchor. NotebookLM will try to imitate the colors and layout style.
4. Use clear filenames, such as Brand_Guidelines_2026.pdf, not final_v3_really_final.pdf.

Style Control Prompt

Paste this into the Studio panel’s Custom prompt field:

只能使用上傳的來源內容，製作一份專業簡報。
目標受眾：[受眾描述]
格式：Presenter Slides（文字少、視覺比重大）
頁數：10-12 頁（每頁一個核心訊息）
每頁：醒目標題最多 8 個字，最多 3 個 bullet，每個 bullet 不超過 12 個字。
設計：低飽和專業色調、乾淨的 sans-serif 字體、大量留白。
視覺請使用彩色鉛筆 hand-drawn illustration 風格。
禁止使用漸層色、霓虹色、發光效果。
背景使用淺色暖色調。

“Colored pencil style” is a repeatedly tested instruction that reduces the AI look. It makes generated illustrations look hand-drawn instead of synthetic. NotebookLM style control is not stable; the same prompt can produce different results, so a few tries may be needed.

CJK Text Rendering Problem

This is NotebookLM’s biggest pain point today. Slides are essentially AI-generated images, and CJK text rendering often distorts. The most practical workaround: use NotebookLM to generate outline and structure, but finish the final deck in Google Slides with Gemini side panel. This bypasses the garbled text problem entirely.

5 Concrete Tips to Avoid the AI Look

1. Colors: Cut Gradients and Specify Hex Codes

AI’s favorite “blue-purple gradient + dark background” is the biggest visual trap. Write exact hex colors into the prompt:

配色：#1A365D（主色）+ #F7FAFC（背景）+ #38B2AC（強調色）。
只用純色塊。不要漸層、不要霓虹、不要彩虹配色。

Low-saturation palettes, such as navy + off-white + one accent color, work better than vague descriptions like “professional colors.”

2. Fonts: Use What the Audience Recognizes

For CJK audiences, use Noto Sans or an appropriate local sans-serif font. Gamma supports uploading custom fonts (.ttf / .otf). After uploading, enter “Switch to uploaded custom font” in AI Agent to apply it globally.

3. Images: Use Gemini Instead of Default Assets

The default illustrations from AI presentation tools usually look very AI. A practical workflow is to generate images separately in Gemini, where quality is high and prompt control is fast, or replace them with your own photos. For precise logos, manual work in Figma is still safer.

4. Information Density: One Idea, Three Points

This is the most commonly ignored part. Add this to the prompt:

請嚴格遵守 One-Idea-Per-Slide。若某個章節有 3 個重點，
就拆成 3 頁。每頁正文最多 20 個字。
所有補充細節都移到 speaker notes。

It is better to add ten slides than to cram three ideas onto one.

5. Speaker Notes: Let AI Write the Talk Track

After the deck is generated, paste the content back into Claude and ask it to write speaker notes:

針對以下每張投影片，寫一段 speaker notes（60-100 字）。
包含：開場過場句、這頁的核心論點用口語解釋、
一個具體的例子或數據佐證。
語氣：自然口語，像一個有自信的講者在對同事說話。

When presenting, you do not need to memorize a script; one glance at the notes tells you what to say on each slide.

Outline First, Visualize Later

Many people skip the outline discussion and ask Gamma or NotebookLM to generate a full deck immediately. That produces many similar-looking slides. Spend 10 minutes with Claude or ChatGPT to lock the outline first, then enter the presentation tool.

想做一份關於 [主題] 的簡報，受眾是 [描述]，目標是 [說服/教學/匯報]。
請整理成 10 張投影片的大綱，每張包含：
1. 結論式標題（例如「XXX 超標 12%，三個推手」，避免「XXX 分析」這種空泛標題）
2. 三個支持論點，各不超過 15 字
3. 建議搭配什麼圖表或視覺元素

After you have the outline, paste it into Gamma’s prompt field or save it as a text file and upload it to NotebookLM. Decks made this way have much stronger narrative logic than one-sentence prompts.

This works for school reports too. Organize references and notes into an outline first, then let AI handle layout and visualization. You keep control of the content, and teachers do not need to worry that the whole deck was blindly generated.

FAQ

How do I make AI presentations look less AI-made?

Use a three-step workflow: discuss the outline with Claude or ChatGPT, confirm logic and information filtering, then generate slides in Gamma or NotebookLM, and finally manually adjust colors, fonts, and images. Keep each slide to one core message and move details into speaker notes.

What is the best way to write a Gamma prompt?

Four essentials: audience description, clear page count and section structure, bullet count limits per slide, and color hex codes plus a negative prompt excluding gradients, neon, and AI-style images. Write design instructions in English, then specify the output language at the end.

Can NotebookLM presentation style be controlled?

Custom prompts can specify a rough direction such as “colored pencil style, no gradients,” but control is less stable than Gamma and each generation can differ. A more reliable method is uploading a brand guide PDF as a visual reference so NotebookLM can imitate existing style.

Should design prompts be written in English?

For design control, the difference is obvious. English is more precise for layout parameters such as whitespace, color, and typography, because much of the design training data is English. The practical rule: write design instructions in English, specify output language separately.

---

Penchan’s Take

“Colored pencil style + no gradients” is the most effective phrase I have tested for reducing NotebookLM’s AI look. The reason is simple: AI defaults to high saturation, gradients, and 3D textures that are recognizable immediately. Colored pencil style brings hand-drawn texture and irregularity, so it falls into the AI look less often.

The practical tool split: Claude handles outline thinking because instruction following and logic are strong; Gamma or NotebookLM handles layout generation because they are fast and visually competent; Google Slides or Canva does the final pixel-perfect control. Let each tool do what it is best at, instead of forcing one tool to do everything.

Using English for design instructions came from repeated failures. Localized terms like “minimal style” can map poorly inside the model, while minimal, flat design, extensive whitespace is more stable. Design descriptions in English plus target-language output is the most reliable combination today.

For final reports or formal proposals, NotebookLM is also a fixed path: clean text references first → upload as sources → control style with Custom prompt → generate outline in NotebookLM → finish in Google Slides. The content the audience sees is grounded, and the visuals are manually crafted.

Further Reading

• 2026 AI Presentation Tools: Gamma, NotebookLM, Copilot Compared
• Gamma AI Presentation Tutorial: Build Professional Slides for Free
• NotebookLM Presentation Tutorial: Generate Slides from Documents

---

— Penchan