Vivi Chen：用 AI 解鎖我的跨域新篇章

1. In one sentence

A salesperson from a humanities background who writes no code at all treats AI as a thinking partner to iterate with rather than a wishing well, going within a year from a LINE bot built in two hours to snowballing all the way to independently producing a project management system with a dashboard.

2. The session

Vivi introduces herself as sales director at Ad2iction (艾迪英特股份有限公司), with a career background in media advertising and a purely humanities education. On verification, her LinkedIn positions her beyond sales director as an AI applications consultant and generative AI instructor, consistent with the role she shows here. She said that a year ago even the generative AI conference felt remote, until a small generative AI meetup happened to be held near her office last July or August, she got in by pretending to be a volunteer, and that was her way in. She has since built dozens of small tools and applications, but stresses she still cannot read code and has yet to get her head around Excel's IF and ELSE functions.

She splits the whole journey into three stages, describing each as the same spiral staircase: find the pain point, try to solve it, learn and reflect, adjust the approach, then stack another turn on top.

Stage one: a LINE bot in two hours, and the first pothole

She wired a chatbot called "AI 科技麻瓜好朋友" into LINE OA, with an LLM such as ChatGPT or Claude behind it, plus Cursor and Render's free cloud deployment — two hours from start to live. She also mentioned wiring in a Q&A AI service so people could ask about current events (product name unverified).

But she then tried to repeat the trick, throwing datasets A, B and C at the AI with "integrate these for me", and it took *longer* than building the bot — two or three days of back-and-forth corrections, with an unsatisfying result. From that setback she distilled three lessons:

• Do not treat AI as a wishing well; treat it as a thinking partner. Discuss and clarify the real requirement first, then ask it to help decompose the data, rather than throwing one sentence at it and expecting end-to-end delivery.
• AI is not a shortcut for avoiding a mess; it means facing the core of the problem. If the data itself is a mess, it was already not okay when you were organising it by hand, and you cannot expect AI to turn a poor input into an ideal output out of thin air.
• AI right now gets "excited" very easily and just keeps producing code; she deliberately calls a halt and insists on discussing before writing.

Stage two: simplifying repetitive annoyances

She gave three simplification cases, all driven by Google Apps Script, written by first asking ChatGPT or Claude "I have a Google Sheet, work out how to generate the Apps Script code for it":

• An internal posting helper: internal emails have a fixed but fiddly format (client name and project name have to go inside brackets), simple information with tedious formatting. An email she would have put off for five hours became: pick client, project and recipient from Google Sheet dropdowns, send with one click, done in a minute.
• Ad creative checks: a client's ad creative goes out to Facebook, YouTube, Google Ads, LINE and the company's own Ad2 platform, and the usual problems are wrong dimensions (300×250 made as 301×251) or wrong duration (30 seconds made as 31), taking at least half an hour to an hour to catch by hand. She had the AI write Apps Script to organise the assets in a Google Drive folder into a Google Sheet automatically, listing platform, filename, file type, dimensions and duration in one go — done in a minute.
• A bubble tea survey: too many drink options and too many customisations for office orders, so she photographed the menu, ran OCR on it, and asked the AI to write Apps Script generating a Google Form directly, saving at least half an hour to an hour.

She also noted that different ecosystems have their own comparable automation tools: Apps Script in Google's, Power Automate in Microsoft's, or n8n as introduced in 高見龍's earlier session — any of which handles a repetitive flow in one click. Her lesson from this stage is to start from your own concrete pain point, begin with something simple, build confidence step by step, and act immediately.

Stage three: more varied tools, and bigger potholes

• A Chrome extension: the internal system requires filling in staff performance evaluations, with key performance indicators, resolution rate and satisfaction all needing to be written out by hand — fine for one person, but a lot of repeated copying and pasting for a team of a dozen to thirty. She had the AI write a Chrome extension so that selecting the customer service inbox or meeting notes fields brings four fields straight into a side panel, saving over a hundred copy-pastes by her calculation.
• A travel data scraper: the Tokyo Disneyland site makes you queue to enter and the interface is in Japanese, so she had the AI write a scraper pulling information on dozens of attractions (height restrictions, maps, videos) into Notion, filterable and shareable, saving hundreds of copy-pastes — and later extended it into a live data site.
• A LINE-style screenshot tool: she likes LINE's screenshot feature but uses it to capture web information rather than to send messages, and the Chrome screenshot extensions on the market did not suit her, so she built her own homage. Along the way she wanted to solve the need to drag left and right while capturing, asked the AI, and got three options — fixed width, preserve ratio, adaptive scaling. Adaptive sounded most perfect but was "complex to implement and error-prone". She was briefly ready to accept the adaptive option wholesale, and only after going round in a circle realised all she actually needed was to scale the original frame to seventy or eighty per cent — no such complexity required. Her conclusion: do not accept the AI's proposal wholesale, do not let it pile on features mindlessly, and do not fantasise about getting a perfect solution in one go — do it step by step, one piece at a time.
• A project management system: she built one with a dashboard tracking each sales project's progress (six checkable stages per case). Along the way she found requirements snowballing as development went on, and had to think about things like a dynamic database, which is complex territory for her; judging her own visual sense to be poor and Cursor's output ugly, she switched to generating the screen design with Claude and then feeding Claude's React rendering or code back into Cursor to follow — cross-using different LLMs' strengths to cover her own weaknesses. She finally hit permissions and data security, and came to accept that solo development ultimately needs a professional engineer, describing it as "a tug of war that never stops, forever expanding its own boundary".

3. Figures and cases

• Her first LINE AI Q&A bot, "AI 科技麻瓜好朋友", built within two hours.
• The multi-source data integration case took at least two or three days with a worse result than expected, far longer than the two-hour bot.
• The three Apps Script cases — internal posting, ad creative checks, bubble tea survey — each went from half an hour to over an hour of manual work down to one minute.
• The Chrome extension case: at thirty staff and four fields each, about a hundred and twenty copy-pastes saved.
• The project management system tracks each project across six stages.

4. Lines worth keeping

• "I shouldn't have imagined AI as so wonderful, treating it as a wishing well and just telling it what I wanted; I should have treated it as a thinking partner and had a discussion with it."
• "AI isn't a shortcut for avoiding the mess; it means genuinely facing the core of the problem."
• "Those of us who don't understand code still have to embrace the grey areas of learning the technology — you have to keep moving forward in that peculiar state of understanding a little but not fully."
• "It's a tug of war that can never stop, forever expanding its own boundary."

5. Tools and terms mentioned

LINE OA, ChatGPT, Claude, Cursor, Render, Google Sheets, Google Apps Script, Google Drive, Notion, OCR, Chrome extensions, Power Automate, n8n.

6. Wider observations

Of the three "non-engineers coding with AI" sessions, Vivi's is the most like carrying a hammer around looking for nails: from the excitement of a chatbot in two hours to the frustration of the multi-source data pothole. The lessons she ends up with — do not treat AI as a wishing well, do not let it pile on features mindlessly, do not expect perfection in one shot — are a variation on Peggy Lo's later emphasis on splitting things up, testing and communicating incrementally, except that Vivi stumbled her way there alone while Peggy organised it more systematically into a methodology. Her point about a poor visual sense and cross-using different LLMs to compensate also echoes 海馬's insistence on giving precise colour codes rather than just images when building his brand site — all three speakers are really saying the same thing: how precisely you communicate with AI determines the quality of what it produces.

7. Sources