好的，这是您要求的英文内容翻译成繁體中文的版本。
<code-segment id="0" lang="text">
i'm spinning up a US-targeted soccer clipping account
for the 2026 World Cup window. give me 5 angle ideas.
</code-segment>
→ Human reading (LLM prompt). Translate body. Output:
<code-segment id="0" lang="text">
我要做一個面向美國市場的足球剪輯帳號，
針對 2026 世界盃窗口期。給我 5 個切入角度。
</code-segment>
<code-segment id="1" lang="text">
warmup window: may 6 to may 31 (3-4 weeks)
posting cadence: 2-3 clips per day
</code-segment>
→ Machine reading (config table). Return body verbatim:
<code-segment id="1" lang="text">
warmup window: may 6 to may 31 (3-4 weeks)
posting cadence: 2-3 clips per day
</code-segment>
<code-segment id="2" lang="json">
{
"task": "Summarize the article in 100 words",
"tone": "casual"
}
</code-segment>
→ JSON-shaped prompt (semantically natural language). Translate string values:
<code-segment id="2" lang="json">
{
"task": "用 100 字總結這篇文章",
"tone": "隨意"
}
</code-segment>
<code-segment id="3" lang="python">
Calculate the sum
total = a + b
</code-segment>
→ Mixed (code + comment). Translate ONLY the comment:
<code-segment id="3" lang="python">
計算總和
total = a + b
</code-segment>
The biggest misconception in AI monetization today is the idea that you can make money by 'creating amazing prompts,' 'writing mass articles with AI,' or 'wrapping AI tools.' That's outdated. What the sophisticated crowd overseas is looking at is not prompts but workflows, not mass production but delivery speed, and not AI tools but business units that can be delegated to AI.
This trend is being accelerated by the latest generation of models from ChatGPT and the long-duration agent models from Claude. OpenAI has deployed GPT-5.6 in three sizes—Sol, Terra, and Luna—available via ChatGPT Work, Codex, and API. Sol is positioned as a high-end model for complex coding, knowledge work, research, computer operation, and design, while Terra and Luna are designed for speed and cost efficiency. Meanwhile, Anthropic's Claude Fable 5 is marketed as a model for long-duration, high-difficulty, multi-stage tasks, available through Claude Code, Claude Cowork, and API.
The conclusion is that the monetization strategy currently targeted by overseas AI geeks is 'dual-wielding': creating quickly with GPT and refining deeply with Claude. GPT is strong in cost efficiency, UI, implementation, documentation, and mass production. Claude is strong in long context, complex codebases, persistent understanding of specifications, and self-verification. By combining these, an individual can replicate parts of a production company, research firm, development house, or business improvement consultancy with a very small team.
I've summarized the setup and hardcore usage techniques in a PDF.
If you want it, you can get it from here! 👇
[https://x.com/MakeAI_CEO/status/2027682940847898770?s=20](https://x.com/MakeAI_CEO/status/2027682940847898770?s=20)
Why 'Dual-Wielding' Instead of Just ChatGPT or Claude?
People who fail at AI monetization immediately want to decide 'which AI is the strongest.' However, the strongest users overseas don't treat models like a religion. They treat models as components for specific roles.
OpenAI's GPT-5.6 has announced API pricing of $5 per 1M input tokens and $30 per 1M output tokens for Sol, $2.5/$15 for Terra, and $1/$6 for Luna. Furthermore, GPT-5.6 and later introduced explicit prompt caching and cache retention for over 30 minutes, making it easier to manage costs for repetitive business prompts. In short, it's suited for mass production, repetition, and templated work.
Claude Fable 5 has API pricing of $10 per 1M input tokens and $50 per 1M output tokens. While more expensive than OpenAI's Sol, Fable 5 is described as an agent capable of 'working over several days,' 'planning multiple stages, delegating to sub-agents, and checking its own work.' This makes it suitable for the final polish of high-ticket projects, design reviews, long-context understanding, refactoring, and finding contradictions in specifications.
Furthermore, independent benchmark Artificial Analysis reports that GPT-5.6 Sol achieves intelligence scores close to Claude Fable 5 at a lower evaluation cost, and Sol ranked high in the Codex environment on the Coding Agent Index. What this shows is not 'who won,' but a design philosophy where you should separate the layer that runs cheaply and quickly from the layer that refines at a higher cost.
What really matters in monetization is not the quality of a single output. It's the gross profit. Making money with AI isn't about buying and selling model intelligence; it's about converting the difference in model capability and cost into deliverables that customers can understand.
The Basic Overseas Geek Style: 'Dividing Tasks by Model'
For example, suppose you have a project to build a Web service MVP. A common AI use case in Japan might stop at 'having ChatGPT write code' or 'having Claude find bugs.' Overseas AI geeks decompose it further.
First, use GPT to create market research, LP structure, UI drafts, component design, initial code, and even demo video scripts all at once. GPT-5.6 is described in OpenAI developer materials as having enhanced frontend layout, visual hierarchy, and design judgment. Then, have Claude Fable 5 review it for specification contradictions, code structure, security oversights, and areas likely to break in long-term operation. Fable 5 is described as being for ambitious coding projects, large-scale migrations, complex implementations, and multi-day autonomous sessions.
The key here is not to use AI as a 'human replacement' but to assign AIs to different job roles.
GPT is the product designer and junior implementer who gives shape to ideas. Claude is the senior engineer and reviewer who is picky about specifications. The human is the producer who listens to the customer's problems, decides the goal, and takes responsibility for the deliverables. With this three-layer structure, an individual can operate like a small production company.
Monetization Method 1: Turning AI Prototypes into 'Validation Packs' instead of 'Deliverables'
The easiest way to start is AI prototype production. However, selling 'I'll make a Web app with AI' leads to price competition. The strong way to sell overseas is to make it a package for validating business hypotheses, rather than just making the prototype.
For example, create a product like this for sole proprietors or small SaaS companies:
'I will create an LP, a simple Web app, usage scenarios, price testing copy, and a demo video script in 48 hours. The deliverables will be ready for actual customer interviews or ad testing.'
In this case, GPT handles market comparison, LP wording, UI, initial code, slides, and ad copy. Claude reviews target contradictions, pricing weaknesses, onboarding gaps, code failures, and paths where customers might drop off.
This product is strong because the customer doesn't just 'want code'; they want to 'know quickly if this idea is worth pursuing.' In other words, what you are selling is not AI production, but shortening decision-making.
The price can be designed as a 'Validation Pack for $1,000,' an 'Investor Demo Pack for $2,000,' or an 'Internal Approval PoC Pack for $3,500,' rather than '$200 per page.' The customer is buying materials for internal meetings, sales, fundraising, and ad testing, not labor hours.
Monetization Method 2: Creating 'Deep Specs' with Claude and 'Selling Presentation' with GPT
A particularly strong combination for ChatGPT and Claude is the specification business. It's unglamorous but very solid.
In the AI era, what will increase is not outsourcing finished products, but 'specifications intended for AI to build.' As no-code, AI coding, and internal automation spread, companies lose track of 'what to build,' 'how to explain it,' and 'how much to leave to AI.'
What sells here are PRDs for AI implementation, requirement definitions, user stories, acceptance criteria, screen transitions, test perspectives, and risk lists.
Claude Fable 5 is described as a model suited for long context and complex knowledge work, supporting the understanding of diagrams, tables, and charts in PDFs. Therefore, it's suited for reading meeting minutes, existing materials, spreadsheets, competitor sites, and past failure cases provided by the customer to distill them into deep specifications.
Meanwhile, GPT-5.6 is being deployed via ChatGPT Work to gather context from team tools and files and convert them into deliverables like documents, spreadsheets, and slides. In other words, you take the deep requirements structured by Claude and expand them into internal proposals, sales materials, LPs, emails, and ad copy with GPT. This flow is powerful.
Product names could be 'AI Pre-development Spec Pack,' 'Claude-Reviewed PRD,' or 'Requirement Definition Kit for AI Coding.' Customers buy blueprints that won't fall apart even if thrown at AI or outsourcers. This is a field where it's easy to raise unit prices in B2B.
Monetization Method 3: Selling Improved Traffic via AI Search and AI Referrals
The next growth area is AI Search Optimization. This isn't just traditional SEO, but designing how you get picked up by AI answers like ChatGPT, Claude, Gemini, and Perplexity.
A June 2026 study by SE Ranking found that traffic from AI search engines to websites grew 16x from 2024 to 2026, accounting for 0.32% of all web traffic by 2026. It's still small, but ChatGPT accounts for 74.78% of AI referral traffic, and Claude is showing high growth. Search Engine Journal also noted that Claude's referral traffic grew significantly from January to April 2026, the highest growth rate among target platforms.
What overseas geeks are looking at here is not 'whether AI search will replace SEO.' They are aiming to productize new traffic channels that companies aren't measuring yet.
The sales pitch is this:
'I will investigate which competitors are recommended when your company name, product name, or category is asked in ChatGPT and Claude. I will create FAQs, comparison pages, case studies, and structured descriptions that are easily cited in AI answers. I will re-measure in one month.'
This is not technically too difficult. Use GPT to generate mass search intent patterns, Claude to classify response trends, GPT to create improvement content, and Claude to check for reliability and missing information. You sell it to the customer as a 'measure to increase the probability of being introduced by AI.'
The point in this area is not to hide that AI search traffic is still small. Rather, saying 'It's small now, but it's growing. That's why we should build the measurement foundation and comparison pages now' builds more trust. AI monetization fades for those who exaggerate.
Monetization Method 4: Making Claude Code Sub-Agents the 'Product'
What hardcore overseas developers are doing is building Claude Code sub-agents and MCP integrations as templates for specific tasks.
Claude Code is described as an agentic coding tool that can read codebases, edit files, run commands, and integrate with development tools. It's designed to connect with external tools, databases, APIs, issue management, monitoring tools, Figma, Slack, etc., via MCP. Furthermore, Claude Code has a mechanism to create sub-agents specialized for specific tasks, each with its own context, system prompt, and tool permissions.
What sells here is not just a collection of prompts. It's the AI team configuration for each business task.
For example, you can create templates like these:
'Claude Code Agent Pack for SaaS Maintenance Teams'
Roles: Code reviewer, bug reproduction specialist, DB query checker, test creator, release note creator.
Customers can semi-automate weekly maintenance tasks just by putting this in their repository.
'Improvement Agent Pack for Shopify Operators'
Roles: Product description improver, SEO FAQ generator, review analyzer, inventory data summarizer, campaign LP editor.
By connecting to store data or spreadsheets via MCP, you can have the AI act as an 'assistant store manager.'
'Document Inspection Agent Pack for Professionals'
Roles: Contract summarizer, gap checker, terminology unifier, comparison table creator, client explanation generator.
Use Claude's long context and document understanding, and convert to proposals or emails on the GPT side.
This product is better sold with initial setup included rather than as a standalone template. This is because many customers don't understand the concepts of MCP or sub-agents. Overseas geeks turn this into an 'implementation service with education.' In other words, what they are selling is not a file, but the initial construction to house AI in the business.
Monetization Method 5: Mass Producing with Cheap GPT Models and Performing 'High-Value Judgment' Only with Fable
In AI monetization, those who throw everything at the top-tier model lose. Successful people overseas always differentiate model usage.
GPT-5.6 has Sol, Terra, and Luna models with different price points. OpenAI's developer guide for GPT-5.6 recommends testing by lowering the reasoning level depending on the task after updating quality and efficiency standards for complex production workflows, and limiting 'max reasoning' to difficult, quality-priority tasks rather than using it for everything.
This thinking can be used directly for monetization.
For example, suppose you sell an ad creative improvement service. The work of creating 100 ad copy drafts, 20 types of LP headlines, and 10 short video scripts is handled by the low-cost GPT configuration. From there, the 10 drafts with the most potential are passed to Claude Fable 5 to strictly review target psychology, differentiation from competitors, legal/expression risks, and brand tone. Finally, it's formatted as delivery material with GPT.
From the customer's perspective, it's not 'I made 100 copies with AI,' but 'I generated 100 drafts, selected and improved the top 10 with a high-performance model, and added a test design.' This allows you to raise the price.
In short, use Fable for judgment, not mass production. GPT for mass production and formatting, Claude for selection and deep diving. This division of labor allows you to increase the persuasiveness of the deliverables while keeping AI costs down.
Monetization Method 6: Turning AI Business Improvement into 'Weekly Report Delegation' instead of 'Automation'
What's easy to sell to companies is not flashy full automation. Rather, the first thing that sells is delegating the tedious weekly reporting work.
For sales teams, summarizing meeting notes, CRM, emails, minutes, and reasons for lost deals. For EC, summarizing sales, inventory, ads, reviews, and competitor prices. For recruitment, summarizing applicants, interview notes, progress by job type, and candidate evaluations. These jobs exist in many companies, and employees are doing them every week with a sigh.
ChatGPT Work highlights uses like gathering context from team tools and converting them into spreadsheets, documents, and slides. Claude Code and the Claude Agent SDK provide ways to handle agents that include file reading, command execution, web search, and code editing from Python or TypeScript.
Using these, you can sell 'I will deliver a two-page A4 report for management every Monday morning' instead of 'I will develop a full automation tool.' It can be semi-manual at first. Assuming a human does the final check, let the AI handle aggregation, summarization, anomaly detection, insight generation, and slide creation.
Even at $500/month, 10 companies make $5,000. At $1,500/month, 5 companies make $7,500. Moreover, because the deliverable is clear, it's easy for customers to continue.
The key in this area is not to say 'I can do anything with AI.' Rather, narrow it down to 'I will look only at this data in this format every week.' The more you narrow it down, the more stable the prompts and workflows become, caching and templates work better, and gross profit increases.
Monetization Method 7: Becoming an AI Era 'Reviewer'
Reviewing AI-generated content is surprisingly lucrative. As AI increases, so do people who don't know if they can trust what AI has made. There is a demand for 'reviewers' here.
For example, reviewing an LP made with AI. Reviewing code made with AI. Reviewing a business plan made with AI. Reviewing a contract draft made with AI. Reviewing sales materials made with AI.
At this time, reviewing with humans alone takes time. Reviewing with AI alone makes responsibility ambiguous. So, you combine GPT and Claude.
Have GPT decompose the target and output a wide range of improvement suggestions. Have Claude look for contradictions, omissions, long-context consistency, potential risks, and awkwardness from a customer perspective. Finally, a human selects 'suggestions to adopt' and 'suggestions to ignore' and delivers them as a review report.
Claude Fable 5 emphasizes uses like testing its own work and checking output against goals. If you use this as a selling point, it's more trusted to express it as a 'double review where perspectives are divided across multiple models and the final judgment is made by a human,' rather than 'I reviewed it with Fable.'
Reviewers can start without taking too much delivery responsibility. If you make it 'review only' instead of 'including fixes,' the initial load is light. From there, you can upsell to 'improvement implementation included,' 'monthly review contracts,' or 'creation of internal AI quality control guidelines.'
Common Trait of Those Who Fail: Selling AI as 'Magic'
Reading this far, it might look like a story of 'making $10,000 a month with AI' immediately. However, in reality, more people fail. The reason is simple: they sell AI like magic.
Customers are no longer surprised that AI is amazing. They are surprised when their work finishes early, when materials that lead to sales are created, when internal approval passes, or when weekly hassles decrease.
In other words, what you should sell is not 'AI utilization.' What you should sell is one of the following:
Time reduction.
Shortening decision-making.
Improvement of sales materials.
Automation of internal materials.
Validation before development.
Exposure in AI search.
Review of existing operations.
Quality assurance of code and materials.
AI should be used behind the scenes. In fact, putting it too far in front makes it look cheap. Successful people overseas sell 'I can deliver this result at this speed and price,' not 'I made it with AI.'
Pricing Design: Deciding by 'Human Replacement Value' instead of AI Cost
Beginners decide prices by looking at AI API costs. This is a mistake. Of course, cost management is necessary, but what the customer pays for is not the API cost. The customer pays for the value of reducing outsourcing costs, labor costs, opportunity loss, and decision-making delays.
For example, even if the AI cost for weekly report delegation is $20/month, if it can turn a task the customer spent 5 hours on every week into 1 hour, $500/month is cheap. Even if the AI cost for prototype validation is $100, if it can advance a business judgment the customer worried about for 3 months in 1 week, $3,000 is viable.
However, high-cost models like Claude Fable 5 require careful use. During the promotion period, Fable 5 can be used up to 50% of the weekly usage limit in some paid plans at no additional cost, but after 23:59:59 PT on July 19, 2026, it will not be included in the plan's weekly limit and will require usage credits for continued use, according to Claude's help. API usage is not eligible for the promotion, and standard rates apply.
That's why you shouldn't use Fable for 'everything,' but limit it to 'high-value judgment,' 'final review,' 'long-context understanding,' and 'complex fixes.' Pre-process with cheap GPT configurations and refine with Fable. This is the secret to protecting gross profit.
If You're Actually Starting, This Order is the Most Solid
You don't need to build a SaaS from the start. In fact, those who suddenly build a SaaS are more likely to fail. The smart ones among overseas geeks first sell it as a service, then template it once a repeating pattern is seen, and finally turn it into a tool.
What you should do in the first 30 days is narrow it down to one industry. For example, legal professionals, recruitment agencies, EC operators, B2B SaaS, English conversation schools, real estate, clinics, or production companies. The reason for narrowing the industry is not to increase AI output accuracy, but to make the sales pitch hit harder.
Next, choose one 'tedious task that occurs every week' in that industry. Reports, proposals, minutes, FAQs, comparison tables, ad drafts, LP improvements, code reviews, or customer response analysis. Don't be greedy here.
Then, fix the roles of GPT and Claude. For example, GPT handles the first draft, structure, tables, slides, LP, and implementation. Claude handles review, contradiction detection, long-context understanding, specification organization, and quality checks. The human handles hearing, final judgment, delivery, and improvement proposals.
Finally, make it a monthly subscription rather than a one-time sale. AI utilization is exhausting if it ends as a one-off. It should be a product based on continuity, such as monthly reviews, weekly reports, four improvement proposals per month, once-a-month AI search diagnosis, or twice-a-month prototype improvements.
Concrete Product Examples
To be more specific, you can create products like these now:
1. AI Search Exposure Diagnosis Pack
Investigate the customer's company name, category name, and competitor comparison keywords with ChatGPT and Claude. Visualize which competitors are recommended by AI and improve FAQs, comparison pages, case studies, author profiles, and structured product descriptions. $1,000 for the first time, $500/month for improvements.
2. Pre-AI Development PRD Pack
Define requirements for the customer's idea with Claude and convert them into screen drafts, LPs, and demo materials with GPT. Deliver as specifications that can be handed to AI coding tools or outsourcers. $1,500–$3,000 for the first time.
3. Weekly Management Report Delegation
Summarize sales, ads, inquiries, reviews, and meeting notes every week and deliver as two A4 pages and a few slides. Format with GPT, check for insights and contradictions with Claude. $500–$2,000/month.
4. Claude Code Implementation Starter
Set up CLAUDE.md, sub-agents, review procedures, test procedures, and MCP connection policies for Claude Code in existing repositories. Separate 'tasks left to AI' and 'tasks approved by humans' for the development team. $2,000–$5,000 for initial implementation.
5. AI-Generated Content Review Service
Review LPs, sales materials, code, contract drafts, and business plans made with AI using multiple models, with a human adding final comments. $300–$1,000 per case. Easy to turn into a continuous review contract.
The important thing here is that every product says 'I will shorten the customer's existing business' rather than 'I use AI.' AI is a means. The product is the shortened time and reduced anxiety.
The Winners from Now on are Not 'People Who Can Use AI' but 'People Who Can Delegate Work to AI'
In the early days of AI utilization, people who were good at prompts stood out. Next, people who could write code with AI stood out. But from now on, more unglamorous and strong people will win. Those are the people who can decompose work.
Which task to throw to GPT.
Which task to throw to Claude.
Which task a human should judge.
Which task to template.
Which task to turn into a monthly product.
Which task to use an expensive model for, and which task to use a cheap model for.
People who can design this can make money even if the AI models change. Conversely, those who depend only on small tricks for specific models will disappear in an instant with an update.
The Claude side is aiming high-performance models like Fable 5 at long-duration, high-difficulty tasks. The OpenAI side is expanding GPT-5.6 into ChatGPT Work, Codex, API, multi-agents, and caching, moving closer to the execution side of business. In other words, the battlefield has moved from 'chat response quality' to 'how much of the work can be finished.'
Overseas geeks are already looking there. They are not thinking about 'what to have AI write,' but 'how to team up AI.' Making GPT the production lead, Claude the review lead, and the human the person in charge. This configuration is the shortest route for an individual to work like a small company.
Finally: If You're Entering Now, Don't Aim for Flashy AI Businesses
Translated text (Traditional Chinese (繁體中文) only, no explanations):
目前 AI 變現最大的迷思，就是認為可以透過「創造驚人的提示詞」、「用 AI 大量寫文章」或「包裝 AI 工具」來賺錢。這些想法都已經過時了。海外那些精明的群體關注的不是提示詞，而是 工作流程；不是大量生產，而是 交付速度；不是 AI 工具，而是 可以委派給 AI 的業務單位。
這個趨勢正由 ChatGPT 最新一代的模型和 Claude 的長時效 Agent 模型所加速。OpenAI 已部署了三種尺寸的 GPT-5.6——Sol、Terra 和 Luna——可透過 ChatGPT Work、Codex 和 API 使用。Sol 定位為用於複雜編碼、知識工作、研究、電腦操作和設計的高階模型，而 Terra 和 Luna 則專為速度和成本效益而設計。與此同時，Anthropic 的 Claude Fable 5 則被定位為用於長時間、高難度、多階段任務的模型，可透過 Claude Code、Claude Cowork 和 API 使用。
結論是，目前海外 AI 極客瞄準的變現策略是「雙持」：用 GPT 快速創作，用 Claude 深度打磨。GPT 的優勢在於成本效益、使用者介面、實作、文件撰寫和大量生產。Claude 的優勢在於長上下文、複雜程式碼庫、對規格的持續理解以及自我驗證。透過結合兩者，個人可以用極小的團隊複製出製作公司、研究公司、開發公司或業務改善顧問公司的部分功能。
我已經將設定方式和硬核使用技巧整理成一份 PDF。
如果你想要，可以從這裡取得！👇
https://x.com/MakeAI_CEO/status/2027682940847898770?s=20
為什麼要「雙持」，而不是只用 ChatGPT 或 Claude？
在 AI 變現上失敗的人，會立刻想要決定「哪個 AI 最強」。然而，海外最強的使用者不會把模型當作信仰來崇拜。他們把模型視為 特定角色的元件。
OpenAI 的 GPT-5.6 已公佈 API 定價：Sol 為每 100 萬輸入代幣 5 美元，每 100 萬輸出代幣 30 美元；Terra 為 2.5 美元 / 15 美元；Luna 為 1 美元 / 6 美元。此外，GPT-5.6 及後續版本引入了明確的提示快取和超過 30 分鐘的快取保留，使得管理重複性業務提示的成本更加容易。簡而言之，它適用於大量生產、重複性和模板化的工作。
Claude Fable 5 的 API 定價為每 100 萬輸入代幣 10 美元，每 100 萬輸出代幣 50 美元。雖然比 OpenAI 的 Sol 昂貴，但 Fable 5 被描述為一個能夠「工作數天」、「規劃多個階段、委派給子 Agent 並檢查自己工作」的 Agent。這使其適用於高價專案的最終潤飾、設計審查、長上下文理解、重構以及尋找規格中的矛盾之處。
此外，獨立基準測試機構 Artificial Analysis 報告指出，GPT-5.6 Sol 在較低的評估成本下達到了接近 Claude Fable 5 的智慧分數，並且 Sol 在 Coding Agent Index 的 Codex 環境中排名很高。這顯示的不是「誰贏了」，而是一種設計理念：你應該 將低成本快速運行的層級與高成本精煉的層級分開。
在變現中真正重要的不是單一輸出的品質，而是毛利。用 AI 賺錢不是買賣模型的智慧，而是 將模型能力和成本的差異轉化為客戶能理解的交付成果。
基本的海外極客風格：「按模型分配任務」
例如，假設你有一個建立 Web 服務 MVP 的專案。在日本常見的 AI 使用案例可能停留在「讓 ChatGPT 寫程式碼」或「讓 Claude 找錯誤」。但海外 AI 極客會進一步分解它。
首先，使用 GPT 一次性創建市場研究、LP 結構、UI 草稿、元件設計、初始程式碼，甚至示範影片腳本。OpenAI 開發者資料描述 GPT-5.6 增強了前端佈局、視覺層次和設計判斷力。然後，讓 Claude Fable 5 審查規格矛盾、程式碼結構、安全疏漏以及長期運作中可能崩潰的區域。Fable 5 被描述為用於雄心勃勃的編碼專案、大規模遷移、複雜實作和長達數天的自主工作階段。
這裡的關鍵不是將 AI 用作「人類的替代品」，而是 將 AI 分配到不同的工作角色。
GPT 是將想法具體化的產品設計師和初級實作者。Claude 是對規格挑剔的高級工程師和審查者。人類則是傾聽客戶問題、決定目標並對交付成果負責的製作人。透過這種三層結構，個人可以像一個小型製作公司一樣運作。
變現方法 1：將 AI 原型轉變為「驗證包」，而非「交付物」
最簡單的入門方式是 AI 原型製作。然而，推銷「我會用 AI 做一個 Web 應用程式」會導致價格競爭。海外強而有力的銷售方式是將其包裝成 驗證商業假設的套裝組合，而不僅僅是製作原型。
例如，為獨資經營者或小型 SaaS 公司創建這樣的產品：
「我將在 48 小時內創建一個 LP、一個簡單的 Web 應用程式、使用情境、價格測試文案和一個示範影片腳本。交付成果將可用於實際的客戶訪談或廣告測試。」
在這種情況下，GPT 處理市場比較、LP 措辭、UI、初始程式碼、投影片和廣告文案。Claude 審查目標矛盾、定價弱點、入門引導缺口、程式碼失敗以及客戶可能流失的路徑。
這個產品之所以強大，是因為客戶不只是「想要程式碼」；他們想要「快速知道這個想法是否值得追求」。換句話說，你銷售的不是 AI 製作，而是 縮短決策時間。
價格可以設計為「1,000 美元的驗證包」、「2,000 美元的投資者示範包」或「3,500 美元的內部批准 PoC 包」，而不是「每頁 200 美元」。客戶購買的是用於內部會議、銷售、募資和廣告測試的材料，而不是工時。
變現方法 2：用 Claude 建立「深度規格」，用 GPT 進行「銷售簡報」
ChatGPT 和 Claude 一個特別強大的組合是 規格業務。這看起來不華麗，但非常穩固。
在 AI 時代，增加的將不是成品的外包，而是「旨在讓 AI 建構的規格」。隨著無程式碼、AI 編碼和內部自動化的普及，公司會迷失在「該建構什麼」、「該如何解釋」以及「該留多少給 AI 處理」等問題上。
這裡能賣的是 AI 實作的 PRD、需求定義、用戶故事、驗收標準、畫面轉換、測試觀點和風險清單。
Claude Fable 5 被描述為一款適合長上下文和複雜知識工作的模型，支援理解 PDF 中的圖表、表格和圖表。因此，它適合閱讀客戶提供的會議記錄、現有資料、試算表、競爭對手網站和過去的失敗案例，並將其提煉成深入的規格。
與此同時，GPT-5.6 正透過 ChatGPT Work 部署，從團隊工具和檔案中收集上下文，並將其轉換為文件、試算表和簡報等交付成果。換句話說，你將 Claude 結構化的深度需求，用 GPT 擴展成內部提案、銷售資料、LP、電子郵件和廣告文案。這個流程非常強大。
產品名稱可以是「AI 開發前規格包」、「Claude 審查 PRD」或「AI 編碼需求定義套件」。客戶購買的是即使交給 AI 或外包商也不會出錯的藍圖。這是一個在 B2B 領域容易提高單價的領域。
變現方法 3：透過 AI 搜尋和 AI 推薦銷售改善後的流量
下一個成長領域是 AI 搜尋優化。這不僅是傳統的 SEO，而是設計如何被 ChatGPT、Claude、Gemini 和 Perplexity 等 AI 答案所引用。
SE Ranking 在 2026 年 6 月的一項研究發現，來自 AI 搜尋引擎的網站流量從 2024 年到 2026 年成長了 16 倍，到 2026 年佔所有網路流量的 0.32%。雖然規模仍然很小，但 ChatGPT 佔了 AI 推薦流量的 74.78%，而 Claude 則顯示出高成長。Search Engine Journal 也指出，Claude 的推薦流量在 2026 年 1 月至 4 月間顯著成長，是目標平台中成長率最高的。
海外極客在這裡關注的不是「AI 搜尋是否會取代 SEO」。他們的目標是 將公司尚未衡量的新流量渠道產品化。
銷售話術是這樣的：
「我會調查當 ChatGPT 和 Claude 被問到您的公司名稱、產品名稱或類別時，哪些競爭對手會被推薦。我會創建常見問題集、比較頁面、案例研究和結構化描述，這些內容容易被 AI 答案引用。我會在一個月後重新測量。」
這在技術上並不難。使用 GPT 產生大量的搜尋意圖模式，用 Claude 分類回應趨勢，再用 GPT 創建改善內容，並用 Claude 檢查可靠性和遺漏的資訊。你將其作為「提高被 AI 介紹機率的措施」銷售給客戶。
這個領域的重點不是隱瞞 AI 搜尋流量仍然很小的事實。相反地，說「現在雖然小，但正在成長。所以我們現在就應該建立衡量基礎和比較頁面」，反而能建立更多信任。那些誇大其詞的人，其 AI 變現之路會逐漸黯淡。
變現方法 4：將 Claude Code 子 Agent 變成「產品」
海外硬核開發者正在做的事情是，將 Claude Code 子 Agent 和 MCP 整合建構成特定任務的模板。
Claude Code 被描述為一個具備 Agent 能力的編碼工具，可以讀取程式碼庫、編輯檔案、執行命令並與開發工具整合。它旨在透過 MCP 與外部工具、資料庫、API、問題管理、監控工具、Figma、Slack 等連接。此外，Claude Code 有一個機制可以創建專門用於特定任務的子 Agent，每個子 Agent 都有自己的上下文、系統提示和工具權限。
這裡能賣的不僅僅是提示詞的集合，而是 每個業務任務的 AI 團隊配置。
例如，你可以創建像這樣的模板：
「SaaS 維護團隊用的 Claude Code Agent 包」
角色：程式碼審查者、錯誤重現專家、資料庫查詢檢查員、測試創建者、版本發布說明創建者。
客戶只需將其放入他們的儲存庫中，就能半自動化每週的維護任務。
「Shopify 營運者用的改善 Agent 包」
角色：產品描述改善者、SEO 常見問題產生器、評論分析師、庫存資料摘要者、活動 LP 編輯者。
透過 MCP 連接到商店數據或試算表，就能讓 AI 扮演「助理店長」的角色。
「專業人士用的文件檢查 Agent 包」
角色：合約摘要者、差異檢查員、術語統一者、比較表創建者、客戶說明產生器。
利用 Claude 的長上下文和文件理解能力，並在 GPT 端將其轉換為提案或電子郵件。
這個產品最好以包含初始設定的方式銷售，而不是作為獨立模板。這是因為許多客戶不理解 MCP 或子 Agent 的概念。海外極客將其轉變為「附帶教育的導入服務」。換句話說，他們銷售的不是一個檔案，而是 將 AI 安置在業務中的初始建構。
變現方法 5：用便宜的 GPT 模型大量生產，僅用 Fable 進行「高價值判斷」
在 AI 變現中，把所有事情都丟給頂級模型的人會失敗。海外成功人士總是會區分模型的使用方式。
GPT-5.6 有不同價格點的 Sol、Terra 和 Luna 模型。OpenAI 的 GPT-5.6 開發者指南建議，在更新複雜生產工作流程的品質和效率標準後，應根據任務降低推理層級進行測試，並將「最大推理」限制在困難、品質優先的任務上，而不是用於所有事情。
這個想法可以直接用於變現。
例如，假設你銷售廣告創意改善服務。創建 100 份廣告文案草稿、20 種 LP 標題和 10 個短影片腳本的工作，由低成本 GPT 配置處理。然後，從中選出最具潛力的 10 份草稿，交給 Claude Fable 5 嚴格審查目標心理學、與競爭對手的差異化、法律/表達風險和品牌調性。最後，用 GPT 格式化為交付材料。
從客戶的角度來看，這不是「我用 AI 製作了 100 份文案」，而是「我產生了 100 份草稿，用高效能模型挑選並改進了前 10 名，並加入了測試設計」。這讓你可以提高價格。
簡而言之，將 Fable 用於判斷，而不是大量生產。GPT 用於大量生產和格式化，Claude 用於挑選和深度探索。這種分工讓你能在控制 AI 成本的同時，提高交付成果的說服力。
變現方法 6：將 AI 業務改善變成「週報委派」，而非「自動化」
容易賣給公司的，不是華麗的全自動化。相反地，最先能賣出去的是 委派繁瑣的每週報表工作。
對於銷售團隊：彙整會議記錄、CRM、電子郵件、備忘錄和丟單原因。對於 EC：彙整銷售、庫存、廣告、評論和競爭對手價格。對於招聘：彙整應徵者、面試記錄、按職位類別的進度和候選人評估。這些工作在許多公司都存在，員工們每週都在嘆氣中完成。
ChatGPT Work 強調的用途包括從團隊工具收集上下文，並將其轉換為試算表、文件和簡報。Claude Code 和 Claude Agent SDK 提供了從 Python 或 TypeScript 處理包含檔案讀取、命令執行、網路搜尋和程式碼編輯的 Agent 的方法。
利用這些，你可以銷售「我將在每個週一早上為管理層提供一份兩頁 A4 的報告」，而不是「我將開發一個全自動化工具」。起初可以是半手動的。假設由人類進行最終檢查，讓 AI 處理彙總、摘要、異常檢測、洞察生成和投影片創建。
即使每月 500 美元，10 家公司就是 5,000 美元。每月 1,500 美元，5 家公司就是 7,500 美元。而且，由於交付成果明確，客戶容易繼續使用。
這個領域的關鍵不是說「我可以用 AI 做任何事」。相反地，要縮小範圍，專注於「我每週只會看這個格式的這份數據」。範圍越窄，提示詞和工作流程就越穩定，快取和模板的效果也越好，毛利自然就會增加。
變現方法 7：成為 AI 時代的「審查者」
審查 AI 生成的內容是意想不到的獲利來源。隨著 AI 的普及，也出現了越來越多不知道是否能信任 AI 產出的人。這裡存在著對「審查者」的需求。
例如，審查用 AI 製作的 LP。審查用 AI 製作的程式碼。審查用 AI 製作的商業計劃書。審查用 AI 製作的合約草案。審查用 AI 製作的銷售資料。
在這種情況下，僅靠人類審查耗時費力。僅靠 AI 審查則責任歸屬不明。所以，你需要結合 GPT 和 Claude。
讓 GPT 分解目標並輸出廣泛的改善建議。讓 Claude 從客戶的角度尋找矛盾、遺漏、長上下文一致性、潛在風險和生硬之處。最後，由人類選擇「要採納的建議」和「要忽略的建議」，並將其作為審查報告交付。
Claude Fable 5 強調的用途包括測試自己的工作以及根據目標檢查輸出。如果你將此作為賣點，與其說「我用 Fable 審查過了」，不如將其表述為「一種跨多個模型劃分視角的雙重審查，並由人類做出最終判斷」，這樣更能獲得信任。
審查者可以在承擔較少交付責任的情況下開始。如果你將其設定為「僅審查」，而不是「包含修正」，初始負擔會很輕。從那裡開始，你可以進行追加銷售，例如「包含改善實作」、「每月審查合約」或「建立內部 AI 品質控制指南」。
失敗者的共同特徵：將 AI 當作「魔法」來賣
讀到這裡，可能看起來像是一個「立刻用 AI 月入一萬美元」的故事。然而，實際上失敗的人更多。原因很簡單：他們把 AI 當作魔法來賣。
客戶已經不再對 AI 的驚人之處感到驚訝。他們會感到驚訝的是，當他們的工作提前完成、當能帶來銷售的資料被創建出來、當內部審核通過、或是當每週的麻煩事減少了的時候。
換句話說，你應該銷售的不是「AI 應用」。你應該銷售的是以下其中一項：
時間縮減。
決策縮短。
銷售資料改善。
內部資料自動化。
開發前驗證。
AI 搜尋曝光。
現有業務審查。
程式碼與資料的品質保證。
AI 應該在幕後使用。事實上，把它推得太前面只會讓它看起來廉價。海外成功人士銷售的是「我能以這樣的速度和價格交付這個結果」，而不是「我用了 AI 做出這個」。
定價設計：根據「人類替代價值」而非 AI 成本來決定
初學者會看著 AI API 成本來決定價格。這是個錯誤。當然，成本管理是必要的，但客戶支付的不是 API 成本。客戶支付的是減少外包成本、人事成本、機會損失和決策延遲的價值。
例如，即使週報委派的 AI 成本是每月 20 美元，但如果它能將客戶每週花費 5 小時的任務縮短為 1 小時，那麼每月 500 美元就是便宜的。即使原型驗證的 AI 成本是 100 美元，但如果它能將客戶苦惱了 3 個月的商業判斷在 1 週內推進，那麼 3,000 美元是可行的。
然而，像 Claude Fable 5 這樣的高成本模型需要謹慎使用。根據 Claude 的幫助說明，在促銷期間，Fable 5 可以在某些付費方案中，於每週用量限制內使用最多 50% 而無需額外費用，但在 2026 年 7 月 19 日太平洋時間晚上 11:59:59 之後，它將不再包含在方案的每週限制內，需要消耗使用額度才能繼續使用。API 使用不適用於此促銷活動，並需支付標準費率。
這就是為什麼你不應該將 Fable 用於「所有事情」，而應將其限制在「高價值判斷」、「最終審查」、「長上下文理解」和「複雜修正」上。先用便宜的 GPT 配置進行預處理，再用 Fable 進行精煉。這是保護毛利的祕訣。
如果你真的要開始，這個順序最穩固
你不需要從頭開始建立一個 SaaS。事實上，那些突然建立 SaaS 的人失敗的機率更高。海外極客中的聰明人，會先將其作為服務銷售，一旦看到重複模式就將其模板化，最後才將其變成工具。
你應該在最初的 30 天內做的是，鎖定一個行業。例如，法律專業人士、招聘機構、EC 營運商、B2B SaaS、英語會話學校、房地產、診所或製作公司。鎖定行業的原因不是為了提高 AI 輸出的準確性，而是為了讓銷售話術更有力。
接下來，選擇該行業中一個「每週都會發生的繁瑣任務」。報告、提案、會議記錄、常見問題、比較表、廣告草稿、LP 改善、程式碼審查或客戶回應分析。在這裡不要貪心。
然後，固定 GPT 和 Claude 的角色。例如，GPT 負責初稿、結構、表格、投影片、LP 和實作。Claude 負責審查、矛盾檢測、長上下文理解、規格整理和品質檢查。人類負責傾聽、最終判斷、交付和改善提案。
最後，將其設為每月訂閱，而不是一次性銷售。如果 AI 應用只是一次性的，很快就會耗盡。它應該是一個基於持續性的產品，例如每月審查、每週報告、每月四次改善提案、每月一次 AI 搜尋診斷，或每月兩次原型改善。
具體的產品範例
更具體地說，你現在就可以創建像這樣的產品：
1. AI 搜尋曝光診斷包
用 ChatGPT 和 Claude 調查客戶的公司名稱、類別名稱和競爭對手比較關鍵字。將 AI 推薦了哪些競爭對手視覺化，並改善常見問題集、比較頁面、案例研究、作者簡介和結構化產品描述。首次 1,000 美元，後續改善每月 500 美元。
2. AI 開發前 PRD 包
用 Claude 為客戶的想法定義需求，並用 GPT 將其轉換為畫面草稿、LP 和示範材料。以可交給 AI 編碼工具或外包商的規格形式交付。首次 1,500 至 3,000 美元。
3. 每週管理層報告委派
每週彙整銷售、廣告、詢問、評論和會議記錄，並以兩頁 A4 和幾張投影片的形式交付。用 GPT 格式化，用 Claude 檢查洞察和矛盾。每月 500 至 2,000 美元。
4. Claude Code 導入啟動包
在現有儲存庫中為 Claude Code 設定 CLAUDE.md、子 Agent、審查程序、測試程序和 MCP 連接策略。為開發團隊區分「委派給 AI 的任務」和「由人類批准的任務」。初始導入費用 2,000 至 5,000 美元。
5. AI 生成內容審查服務
使用多個模型審查用 AI 製作的 LP、銷售資料、程式碼、合約草案和商業計劃書，並由人類加入最終評論。每次 300 至 1,000 美元。容易轉變為持續的審查合約。
這裡的重點是，每個產品都是在說「我將縮短客戶的現有業務」，而不是「我使用了 AI」。AI 是手段。產品是縮短的時間和減少的焦慮。
從現在開始的贏家不是「會用 AI 的人」，而是「能將工作委派給 AI 的人」
在 AI 應用的早期，擅長提示詞的人脫穎而出。接下來，能用 AI 寫程式碼的人脫穎而出。但從現在開始，贏家將是那些看起來更樸實但更強大的人。那就是能夠分解工作的人。
哪個任務交給 GPT。
哪個任務交給 Claude。
哪個任務由人類判斷。
哪個任務做成模板。
哪個任務變成每月產品。
哪個任務使用昂貴的模型，哪個任務使用便宜的模型。
能夠設計這些的人，即使 AI 模型改變也能賺錢。相反地，那些只依賴特定模型的小技巧的人，會在版本更新時瞬間消失。
Claude 這邊正將 Fable 5 這類高效能模型瞄準長時間、高難度的任務。OpenAI 這邊則將 GPT-5.6 擴展到 ChatGPT Work、Codex、API、多 Agent 和快取，更貼近業務的執行面。換句話說，戰場已經從「聊天回應品質」轉移到「能完成多少工作」。
海外極客已經看到了這一點。他們思考的不是「要讓 AI 寫什麼」，而是「如何讓 AI 組隊合作」。讓 GPT 擔任生產總監，Claude 擔任審查總監，人類擔任負責人。這個配置是個人像小型公司一樣運作的最短路徑。
最後：如果你現在要入場，不要追求華麗的 AI 業務
最重要的事情是，我想告訴那些現在才開始的人，不要追求花俏的 AI 生意。做 AI 應用、做 AI 媒體、賣 AI 教材、賣 AI 提示詞。當然，這些都有潛力，但競爭也同樣激烈。
更穩健的做法是，把 AI 插入到已經有金錢流動的業務中。報告、銷售資料、規格書、程式碼審查、招募資料、常見問題、競爭者研究、廣告優化、週會資料。工作越不起眼，AI 雙手互搏的效果就越明顯。
結合 ChatGPT 與 Claude 來賺錢的本質，不是「用 AI 輕鬆賺錢」。而是透過模型角色的分工，快速、便宜、持續地清理那些人們覺得繁瑣而拖延的工作。
之所以看起來只有海外極客在這麼做，是因為他們不把 AI 當作「文字生成工具」。對他們來說，AI 是一套用來重新安排生產、實施、審查、研究與品質控管角色的組件。
所以，你應該先學的唯一提示詞是：
「當這項工作被分為五個階段——生成、組織、執行、驗證、交付——請區分哪些階段應該交給 GPT、哪些階段應該交給 Claude、以及哪些階段應該由人類負責。」
每次都能問出這個問題的人，將在下一波 AI 變現中勝出。
