AIPitchPack, paste a URL, get a custom pitch deck in 15 minutesAIPitchPack，貼上網址，15 分鐘拿到客製提案
A solo-built internal AI tool: the prospect's own mascot, content and design, all generated automatically.我獨立打造的內部 AI 工具：客戶的專屬吉祥物、內容和設計，全部自動生成。
01 Background背景
notso.ai offers custom mascot AI agents. For a B2B team, the hardest part of reaching a prospect is showing them the value of the product. We validated that the more a pitch deck was tailored to them, the more often it moved the conversation forward.notso.ai 提供客製化的吉祥物 AI Agent。對一個 B2B 團隊來說，最難的是讓潛在客戶看見產品的價值。我們驗證過，一份提案簡報客製化的程度越高，越能推進到下一次對話。
02 Problem問題
But a genuinely customized deck cost too much time and effort:但一份夠客製的 deck，要耗費的時間和精力都太高：
Each one needed market research, a pain-point read, a matching solution, an early design of the prospect's own mascot, the underlying tech, and ways to collaborate.每份都得有市場調查、痛點判讀、對應解法、潛在客戶專屬吉祥物的初步造型、後端技術、合作方式。
Data was assembled by hand, and output meant hopping between disconnected generation tools, with no efficient workflow.資料要一筆筆手工整理，產出還得在彼此不相連的生成工具間來回切換，沒有一條有效率的工作流。
Each deck took one to two days, while sales juggled everything else.一份 deck 要 1 到 2 天，而 sales 同時還扛著其他工作。
Image models and copy kept drifting, so we spent a lot of time tuning prompts.生圖模型和文案經常跑掉，需要花費大量時間調整 prompt。
The most effective thing we could do was also the most time-expensive.最有效的做法，恰好是時間成本最高的做法。
03 Challenge挑戰
How might we help a busy startup team produce a customized deck in minutes, customized enough that the client feels it was made just for them?我們如何讓一支業務繁忙的新創團隊，在幾分鐘內就做出一份客製簡報，而且客製到讓潛在客戶覺得「這是為我做的」？
04 Research & Insights研究與洞察
I analyzed past decks that earned a reply, decomposing what each needed and where the customized feeling actually came from, then mapped the sales workflow to draw the line between what to automate and what to leave to the user, and finally defined the assets and prompt structure the system had to produce.方法上，我分析過去成功換到回覆的 deck，拆解每份需要哪些元素，以及「客製感」到底從何而來，再盤點 sales 的工作流，畫出哪些該自動化、哪些留給使用者，最後定義系統必須產出的素材和 prompt 結構。
The most deeply customized part, the strategic point of view, is where the user's own input matters most. Light and medium customization can be locked into the prompt and generated automatically, so people spend their judgment only where it counts.客製程度最高的部分（策略觀點），需要使用者親自輸入。輕度到中度的客製，則可以用 prompt 固定下來、交給系統自動生成，人只專注在最需要判斷的地方。
05 Solution解法
In the end the flow split into three stages, ① client info, ② generate mascot and imagery, ③ edit the deck content. AI APIs sit underneath, so each deck auto-adapts its content to the prospect's industry.最終我把流程拆成三個階段：① 填客戶資料 ② 生成吉祥物與圖像 ③ 編輯簡報內容。底層串接 AI API，每份 deck 都會依潛在客戶的產業自動調整內容。
I wrote a brand visual-identity spec (material, proportion tiers, a locked eye formula, mockup style) and routed every mascot generation through one prompt builder, so every output stays consistent.我寫了一份品牌視覺規範（材質、比例 tier、固定的眼睛公式、mockup 風格），並讓每一次吉祥物生成都走同一個 prompt builder，使吉祥物的所有產出維持一致。
06 Design in detail設計細節
1 · Automated research and analysis1 · 資料搜集與分析，自動化
Auto-searches the prospect's web and brand info, and turns it into the market and industry analysis this deck needs.自動搜尋潛在客戶的網路與品牌資訊，整理成這份 deck 需要的市場與產業分析。
2 · One-click mascot media pack2 · 一鍵吉祥物媒體包
One click generates the client's own mascot in many emotions, poses, even holiday looks.一鍵生出客戶專屬吉祥物，含各種情緒、動作，甚至節日造型。
3 · Edit in-system, with AI assist3 · 系統內編輯，AI 輔助
Change the deck right in the editor with a live palette, and let AI rewrite copy down to font size.在編輯器裡直接改簡報、即時換色盤，AI 還能幫忙重新潤稿，連字級大小都能調。
07 Impact影響
A custom deck went from one to two days to about 15 minutes, the bottleneck that made customization unaffordable is gone.一份客製 deck 從 1 到 2 天降到約 15 分鐘，那個讓客製「做不起」的瓶頸消失了。
08 Reflection反思
Running this solo with AI as my build team, the hardest part wasn't deciding the architecture, it was making each tool reliable and keeping manual testing under control.一個人用 AI 當開發團隊，最難的不是決定架構，而是讓每個工具夠可靠，以及控制手動測試的時間。
Every tool wants its prompt structured differently, and “works once” is not “works every time.” Next time I'd split the work across subagents first, each keeping a detailed record on its own Notion page and reporting to a main agent I can query anytime.因為每個工具底下的 prompt 結構都不一樣，而且「能跑一次」跟「每次都能跑」是兩回事。下次我會先把任務拆給 subagent，每個 agent 都在自己的 Notion 頁面留詳細紀錄，回報給一個我隨時能調閱的 main agent。