Building a Vocabulary App with Claude Using Zero Code: From Requirements Description to Deployment

Introduction: An Era Where Non-Programmers Can Build Apps

A user learning Brazilian Portuguese developed a fully functional vocabulary app simply by conversing with Claude AI—without writing a single line of code. The app features human-like pronunciation, verb conjugation lookup across 14 tenses, SM2 spaced repetition review, and works on both mobile and desktop.

This case proves once again: in the age of AI programming, people who can describe requirements well are more valuable than people who can write code.

Detailed App Features

Intelligent Word Lookup

The core feature of this Brazilian Portuguese dictionary app is intelligent word lookup. Users don't need to know the infinitive form of a word—regardless of which conjugated tense form they input, the system automatically identifies and finds the verb's base form. This is extremely practical for language learners, since what you encounter in actual reading is usually the conjugated form.

Query results include:

• Word definitions and human-like pronunciation
• If it's a verb, a complete conjugation table across all 14 tenses
• Conjugation pattern explanations (e.g., present indicative ending change rules)
• Identification of the specific tense and person of the input word
• Usage explanations and example sentences for that tense (each with a pronunciation button)
• Synonym expansion

Verb Conjugation System

Verb conjugation is one of the major challenges in Portuguese learning. The Portuguese verb conjugation system is among the most complex in the Romance language family—a single verb must be conjugated according to tense (present, past, future, etc.), mood (indicative, subjunctive, imperative, etc.), and person (first/second/third person, singular and plural). While Brazilian Portuguese simplifies some person usage in spoken language (e.g., less frequent use of "tu" and "vós"), written and formal contexts still retain the complete conjugation system. Fourteen tenses means a regular verb has dozens of conjugated forms, and irregular verbs (such as high-frequency words like ser, ir, ter) don't follow any patterns at all and must be memorized individually.

The app first displays general patterns, then lists specific conjugation forms for four persons. The first time you look up a particular verb, it takes about 10 seconds to load (because it generates complete conjugation data for 14 tenses in real-time via the OpenAI API), but subsequent lookups of the same word load instantly—indicating the system has caching optimization, storing API results locally or in a server-side database to avoid redundant calls and extra costs.

SM2 Spaced Repetition Memory System

Users can add queried words to their word bank, and the system schedules reviews based on the SM2 spaced repetition algorithm. SM2 (SuperMemo 2) is a spaced repetition algorithm proposed by Polish researcher Piotr Wozniak in 1987 and serves as the theoretical foundation of modern memory software like Anki. Its core principle is based on the Ebbinghaus forgetting curve: human memory decays exponentially over time, but reviewing at the critical point just before forgetting significantly extends memory retention. The algorithm maintains an "easiness factor" (EF value) for each learning item and dynamically adjusts the interval until the next review based on the user's recall quality score—high scores lengthen the interval, low scores shorten it or restart the cycle, achieving personalized optimal review scheduling.

The review mode is designed flexibly: sometimes it shows a Portuguese word for you to recall the meaning, sometimes it plays pronunciation for you to identify, and sometimes it gives you Chinese and asks you to recall the Portuguese—this diversified testing approach helps build stronger memory connections. Cognitive science research shows that retrieving the same knowledge from different angles (i.e., "interleaved practice") promotes long-term memory formation more effectively than single-method repetition.

There's also a check-in weekly report feature, with learning records retained for 90 days.

Technical Architecture and Running Costs

Services Used

The entire app involves the following services:

1. Claude AI (Max plan): Responsible for all code generation throughout the development process. Claude Code is Anthropic's programming assistant tool that allows users to generate, debug, and deploy code through natural language conversation. Unlike traditional IDE plugins (such as GitHub Copilot), it emphasizes end-to-end project building capability—from architecture design to implementation to deployment, all completed in conversational form. The Max plan provides a longer context window and stronger reasoning capabilities, suitable for handling complex tasks at the full project level. The author mentioned that after subscribing to Claude, they cancelled many other tools, considering it excellent value.

2. OpenAI API (paid): Used for word lookup, definition generation, and other intelligent features. This leverages the multilingual understanding capabilities of large language models—GPT-series models include substantial Portuguese language data in their training corpus, enabling accurate generation of definitions, conjugation forms, and example sentences. By calling the API rather than using the ChatGPT interface directly, developers can seamlessly integrate AI capabilities into their own application logic.

3. ElevenLabs API (paid): Provides human-level speech synthesis. ElevenLabs uses deep learning models to generate near-human-quality voice output, capturing subtle features like intonation, rhythm, and emotion. Unlike the mechanical feel of traditional TTS systems, this high-quality voice is especially critical for language learning—learners need to hear natural pronunciation to build correct phonological memory. Users can select their preferred voice through a voice ID, ensuring consistency in voice style throughout the learning process.

4. Free cloud deployment services: The current cloud deployment ecosystem offers multiple free options friendly to personal projects, such as Vercel, Netlify (frontend hosting) and Railway, Render (backend services). These platforms typically use a "free tier + pay-as-you-go" model, and for low-traffic applications like personal learning tools, the free tier is more than sufficient. Deployment only requires pushing code to a Git repository and connecting to the platform for automatic completion—Claude guides you through all configuration.

The actual running cost is only the fees for the OpenAI and ElevenLabs APIs; all other infrastructure is free.

Four Steps from Idea to Launch

1. Register accounts: You only need to register for Claude AI first; it will progressively tell you about other services needed
2. Describe your requirements: Describe the features you want in detail using natural language
3. Iterate and test: Provide feedback on generated results, gradually refining
4. Deploy to the cloud: Claude provides every terminal command and expected result step by step

Core Methodology for Communicating with AI

Describe Requirements Like You're Talking to a Product Manager

The author emphasizes that AI can't understand vague descriptions like "make a word app, make it look nice, with sound." The correct approach is to communicate with AI as if it were a product manager:

"When I'm learning Portuguese, I often forget verb infinitives. I want to look up word meanings, and if it's a verb, I want all tense conjugations; if it's a noun, I want gender markings."

If you're worried about unclear descriptions, give specific examples: "If I give you a conjugated form in some tense, then I want to see such-and-such in the results."

This communication approach is essentially "Requirements Engineering"—the most critical yet most easily overlooked phase in software development. In traditional software projects, rework costs due to unclear requirements account for 40%-60% of total development costs. While the AI programming era has dramatically reduced coding costs, the quality of requirement expression still directly determines the quality of the final product.

The More Detailed Your Description, the Higher the Fidelity

The author's experience is: if AI doesn't fully reproduce your requirements, it's not necessarily the AI's fault—it's very likely you didn't describe things clearly enough. When you discover issues during testing, continue providing feedback: "This part isn't what I wanted; what I want is..."

This process is similar to "user story refinement" in agile development—first give the general direction, then supplement details and edge cases through continuous dialogue. AI's contextual understanding ability makes this progressive requirement clarification very efficient.

Don't Panic About Bugs: Screenshots Are All You Need

When using Claude Code to build small web apps, bugs are nothing to fear, because the simplest solution is to take a screenshot—send a screenshot of the problematic page directly into the chat, and AI can locate and fix the issue. This leverages the visual understanding capabilities of multimodal AI models—Claude can "see" layout misalignment, style anomalies, or functional failures in interface screenshots, correlating visual information with code logic to quickly pinpoint the root cause.

If You Can't Keep Up, Just Say So

If AI gives too much content too quickly and users with no programming background feel overwhelmed, just tell it directly: "I can't keep up with your pace; you need to tell me step by step." AI will slow down and guide you progressively. This reflects an important characteristic of large language models—they can dynamically adjust the granularity and complexity of output based on user feedback, essentially building a "mental model" of the user's technical level through conversation.

Key Insights from Zero-Code Development

The core insight from this case isn't about the technical implementation itself, but about a shift in mindset:

1. Requirement description ability > Coding ability: AI can write code, but it can't figure out what you want for you. In the AI era, "product thinking" and "requirement expression ability" are becoming scarcer skills than programming syntax.

2. No need to understand the tech stack in advance: You don't need to know where data is stored or what algorithms to use—AI will guide your choices. In traditional development, technology selection alone requires years of accumulated experience, but now AI can recommend the most suitable solution based on your specific scenario.

3. Iterative thinking matters: Don't expect perfection on the first try; gradually approach the ideal result through continuous feedback. This aligns with "prototyping" and "agile development" philosophies in software engineering—rapidly produce a minimum viable version, then continuously optimize.

4. The barrier to personalized tools has disappeared: Anyone can customize tools for their specific learning needs. This means "long-tail demands" finally have economically viable solutions—niche tools with user bases too small for commercial companies to develop can now be created by anyone.

This methodology is completely transferable to other language learning scenarios—whether you're studying French, Japanese, or Arabic, just replace "Portuguese" with your target language, describe your specific needs, and you can build your own learning tool.

Key Takeaways

• Through conversation with Claude AI, you can develop a fully functional vocabulary app without any programming background, including human-like pronunciation, verb conjugation lookup, and spaced repetition review
• The key to communicating with AI is describing requirements like you would to a product manager—providing specific scenarios and examples rather than vague feature descriptions
• The entire project only requires two paid APIs (OpenAI and ElevenLabs); infrastructure and deployment can all be done for free
• Bugs can be fixed by simply sending screenshots to AI; if you can't keep up, you can ask AI to slow down and guide you step by step
• Core insight: People who can describe requirements are more valuable than people who can write code; in the AI era, the barrier to personalized tool development has disappeared