<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Lian0123 (連永立)</title>
    <description>The latest articles on DEV Community by Lian0123 (連永立) (@lian0123).</description>
    <link>https://dev.to/lian0123</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4012517%2F3085ca60-eee7-47c6-ac98-a9d16fcc1f3b.jpeg</url>
      <title>DEV Community: Lian0123 (連永立)</title>
      <link>https://dev.to/lian0123</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/lian0123"/>
    <language>en</language>
    <item>
      <title>Notes on Introducing AI into Vibe Coding Development (Part 2) — Why Are Tokens Burning So Fast?</title>
      <dc:creator>Lian0123 (連永立)</dc:creator>
      <pubDate>Mon, 26 Oct 2026 15:36:10 +0000</pubDate>
      <link>https://dev.to/lian0123/notes-on-introducing-ai-into-vibe-coding-development-part-2-why-are-tokens-burning-so-fast-2jef</link>
      <guid>https://dev.to/lian0123/notes-on-introducing-ai-into-vibe-coding-development-part-2-why-are-tokens-burning-so-fast-2jef</guid>
      <description>&lt;h1&gt;
  
  
  Introduction
&lt;/h1&gt;

&lt;p&gt;Have you ever run into this situation? You ask AI to fix a few bugs or adjust a screen, and your usage allowance is almost gone before much work has been completed.&lt;/p&gt;

&lt;p&gt;As AI becomes more capable of handling complex development tasks, we naturally start giving it more work. But greater capability does not mean every run is efficient. Sometimes, what really consumes resources is repeatedly searching for information, guessing requirements, heading in the wrong direction, and fixing problems introduced by earlier changes.&lt;/p&gt;

&lt;p&gt;Model choice, reasoning settings, context length, and platform billing policies all affect cost. But among the things developers can control, one question is especially worth addressing first: Are we making AI do too much work that could have been avoided?&lt;/p&gt;

&lt;p&gt;This article covers several common situations that consume tokens, along with working habits I believe are worth developing.&lt;/p&gt;

&lt;h1&gt;
  
  
  1. First, Understand Where the Tokens Go
&lt;/h1&gt;

&lt;p&gt;Tokens are the basic units a model uses to process content. They are not directly equivalent to words or characters. In addition to our instructions and the AI’s responses, code, documents, conversation history, and tool results sent to the model may all become part of its input.&lt;/p&gt;

&lt;p&gt;Some reasoning models also generate reasoning tokens that are not directly shown in the final response. How these are billed or counted against usage limits varies by model and platform.&lt;/p&gt;

&lt;p&gt;From the workflow a user can observe, a development task can roughly be broken down as follows:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Understand the requirements → Find information → Analyze and plan → Implement → Validate → Revise if necessary
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This is a useful way to describe the workflow, not a fixed sequence of internal thinking steps followed by every model.&lt;/p&gt;

&lt;p&gt;If the requirements are vague, AI has to fill in the gaps with assumptions. If it searches in the wrong place, it has to search again. If the acceptance criteria are unclear, it may keep working on things you never needed. Each extra loop can mean another round of model calls, data retrieval, and output.&lt;/p&gt;

&lt;p&gt;Also, API token charges, subscription usage allowances in chat products, and request rate limits are different mechanisms. They cannot be directly converted into one another. Feeling that your allowance is running out quickly is not, by itself, evidence that a provider has reduced the model’s capabilities.&lt;/p&gt;

&lt;h1&gt;
  
  
  2. Unable to Find the Target: Spending Heavily Before Making Any Changes
&lt;/h1&gt;

&lt;p&gt;Suppose you need to change the code for a particular screen. You would usually start by asking: Which page? Which component? Which API? Then you would narrow the scope using search or the project’s directory structure.&lt;/p&gt;

&lt;p&gt;AI needs those clues too.&lt;/p&gt;

&lt;p&gt;A request like “There’s a bug with orders in production. Please fix it” may force an agent to investigate everything from the project structure, routes, screens, and services to the database. Tools and indexes affect how it searches, but insufficient information can certainly broaden the search.&lt;/p&gt;

&lt;p&gt;Remember: AI cannot read your mind. When clues are missing, it can only form assumptions based on the information available—and those assumptions may be wrong.&lt;/p&gt;

&lt;h2&gt;
  
  
  Check the Most Likely Cause First, Then Broaden the Search
&lt;/h2&gt;

&lt;p&gt;Before giving instructions, try to organize the symptoms, reproduction steps, expected and actual results, error messages, and potentially relevant pages, APIs, or files.&lt;/p&gt;

&lt;p&gt;For example, instead of saying “Orders are broken,” you could write:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;After switching to the second page of the order list and clicking Cancel Order, the screen still shows the original status. It only updates after a refresh. Please first check the list’s state updates and cache invalidation logic, confirm the cause, and then make changes.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;This description helps AI prioritize the most likely path.&lt;/p&gt;

&lt;p&gt;However, you must distinguish a possible cause from a confirmed fact. If a frontend caching issue is only your guess, say so explicitly. Otherwise, AI may treat an incorrect premise as an established conclusion. If the first hypothesis does not hold up, let it broaden the search based on evidence.&lt;/p&gt;

&lt;h2&gt;
  
  
  Build an Index So AI Knows Where to Start
&lt;/h2&gt;

&lt;p&gt;As a project grows, you can maintain a concise architecture index covering the main modules, routes, data flows, core services, and test entry points. Where supported, semantic indexes or code knowledge graphs can also help locate relevant symbols.&lt;/p&gt;

&lt;p&gt;The purpose of an index is to provide starting points, not to rewrite the entire project as a second set of documentation. The tradeoff is maintenance: when the code changes, key index entries must change too. Outdated documentation can be more misleading than no documentation at all.&lt;/p&gt;

&lt;p&gt;Human review is still necessary. What you want to maintain is a record of facts, not hallucinations.&lt;/p&gt;

&lt;h2&gt;
  
  
  Do Not Treat Conversation History as the Project’s Only Memory
&lt;/h2&gt;

&lt;p&gt;After a long development session, a single conversation may contain old requirements, failed approaches, temporary decisions, and assumptions that were later rejected. Once the context has been compressed, some details may also be incomplete.&lt;/p&gt;

&lt;p&gt;Record important decisions, the current state, and remaining tasks in documents that can be inspected and updated. When a conversation has drifted away from the current situation, start a new session with a concise handoff summary, and instruct AI to use the current files and verifiable results as the source of truth.&lt;/p&gt;

&lt;p&gt;Starting a new conversation does not automatically restore modified code. If you need to roll back, you still have to inspect the Git diff or the tool’s checkpoints.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fxsxwam9p5aqoz66arieb.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fxsxwam9p5aqoz66arieb.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h1&gt;
  
  
  3. Understanding Documents: Give Each Step to the Right Tool
&lt;/h1&gt;

&lt;p&gt;Another common scenario is asking a development agent to read a specification document and immediately start implementing it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;PDF/Word → Content extraction → Specification analysis → Comparison with the code → Implementation and validation
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This workflow is reasonable, but that does not mean every step needs to be handled by the same expensive model.&lt;/p&gt;

&lt;p&gt;For documents with a text layer, use a dedicated tool to extract the content first. Scanned documents may require OCR—optical character recognition. When tables, flowcharts, or layout relationships matter, preserve their structure and use a model with visual understanding when necessary. Converting everything to plain text can also lose important information.&lt;/p&gt;

&lt;p&gt;Once the content is available, choose a suitable model to organize the specifications based on their complexity. If the developer already understands the domain, directly providing confirmed technical requirements can often eliminate an extra round of interpretation and restatement.&lt;/p&gt;

&lt;p&gt;However, dividing the work is not free. Handoffs between models introduce costs through repeated reading, information transfer, and possible omissions. If a short document can be handled well in one pass, deliberately splitting it among multiple agents may not be worthwhile.&lt;/p&gt;

&lt;p&gt;Also, make the priority of information sources explicit. For example: requirements confirmed for this task take precedence over older specifications; if the document conflicts with the current code, list the differences first and then confirm how to proceed.&lt;/p&gt;

&lt;h1&gt;
  
  
  4. Starting Too Early: The More You Do in the Wrong Direction, the More It Costs
&lt;/h1&gt;

&lt;p&gt;We often send a prompt and wait for AI to deliver a result. Only halfway through—or after everything is finished—do we realize it chose the wrong technology, changed the wrong workflow, or treated test data as the production design.&lt;/p&gt;

&lt;p&gt;Planning first is usually helpful when building a system, adding functionality across modules, or making changes involving data models and permissions.&lt;/p&gt;

&lt;p&gt;For example, GitHub Copilot’s Plan mode can research a task and propose a plan before changing any code. Whether other tools offer the same capability, and what they call it or how it works, depends on the product. Even without a dedicated mode, you can ask AI to propose an approach and identify questions that need clarification before proceeding.&lt;/p&gt;

&lt;p&gt;During planning, it is worth confirming:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Which technologies will be used?&lt;/li&gt;
&lt;li&gt;What outcome is required, and what is outside the scope?&lt;/li&gt;
&lt;li&gt;Which screens, modules, and data will be affected?&lt;/li&gt;
&lt;li&gt;Will the implementation use real data or mocks? Which platforms must it support?&lt;/li&gt;
&lt;li&gt;How will the completed work be accepted?&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Planning also consumes tokens, so every small change does not need to become a full requirements interview. The point is to adjust the amount of upfront clarification to the task’s uncertainty and impact.&lt;/p&gt;

&lt;h1&gt;
  
  
  5. Ambiguous Wording: One Missing Sentence Can Lead to Many Revisions
&lt;/h1&gt;

&lt;p&gt;Sometimes the issue is simple. “Make this match the one above” may feel obvious to a person, but AI may not know exactly what “this” and “the one above” refer to.&lt;/p&gt;

&lt;p&gt;Before sending a request, quickly check four things: &lt;strong&gt;the target, its location, the constraints, and the acceptance criteria.&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Please update the Save button on the member settings page, at the bottom of the form on the right. Use the project’s existing primary button style and preserve the current save logic. For acceptance, confirm that it works properly on both desktop and mobile.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Clarity does not require length. A few sentences that remove key ambiguities are usually more useful than a long description without clear boundaries.&lt;/p&gt;

&lt;h1&gt;
  
  
  6. Cache Misses: You May Pay Repeatedly for the Same Content
&lt;/h1&gt;

&lt;p&gt;Caching can reduce some costs, but first we need to distinguish two things: conversation memory and a prompt cache are not the same concept.&lt;/p&gt;

&lt;p&gt;Prompt caching mainly reuses previously processed input prefixes to reduce repeated computation. It does not simply reuse the previous answer, nor does it mean the model will remember everything or reason more accurately.&lt;/p&gt;

&lt;p&gt;Cache hits typically depend on factors such as matching prefixes, content length, expiration periods, and platform settings. Staying in the same session does not guarantee a hit. Cache read discounts and write charges may also differ across models.&lt;/p&gt;

&lt;h2&gt;
  
  
  A Concrete Cost Example
&lt;/h2&gt;

&lt;p&gt;The following example uses the publicly listed standard prices for the Claude Sonnet 4.6 API, in US dollars per million input tokens. It compares only the unit prices for the same amount of input, excluding output, other tool fees, and special pricing conditions.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9ps8ul07vddd2iq6q9tz.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9ps8ul07vddd2iq6q9tz.png" alt=" " width="800" height="480"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Writing to the cache may cost more; repeated reads are where savings can emerge. For one million fixed input tokens, with one write followed by four reads:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;If every request is billed as regular input, the total is &lt;strong&gt;$15&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;If a five-minute cache is successfully used, the total is &lt;strong&gt;$3.75 + $0.30 × 4 = $4.95&lt;/strong&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This is a simplified calculation assuming the reused content fully qualifies for caching and continues to produce cache hits. It excludes new content added in each turn and all output.&lt;/p&gt;

&lt;p&gt;For API developers who can control request content, it helps to place stable rules, tool definitions, and shared data near the beginning, with frequently changing requirements toward the end. Check usage data to verify actual cache hits. With ready-made development tools, what you can do depends on the controls and visibility the tool provides.&lt;/p&gt;

&lt;p&gt;Accumulating irrelevant history just to pursue cache hits may not be worthwhile either. Cheap unnecessary content is still unnecessary content.&lt;/p&gt;

&lt;h1&gt;
  
  
  7. Misunderstandings: Identify the Level of the Error Before Fixing It
&lt;/h1&gt;

&lt;p&gt;Even with substantial information, AI may still change the wrong screen, choose the wrong component, or design a data structure that does not fit the business. Different situations call for different corrective actions.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Overall Direction Is Wrong: Roll Back and Redefine
&lt;/h2&gt;

&lt;p&gt;If AI misunderstood the requirements from the start and all subsequent changes rest on that incorrect premise, rolling back to the state before the changes and adding clear conditions is usually a cleaner approach.&lt;/p&gt;

&lt;p&gt;First, however, confirm what the tool’s rollback actually covers. Some tools only return to an earlier point in the conversation; others also restore some files. Changes to databases, external services, or changes caused by commands may not be undone.&lt;/p&gt;

&lt;h2&gt;
  
  
  Only One Part Is Wrong: Keep the Results and Narrow the Fix
&lt;/h2&gt;

&lt;p&gt;If you requested three things and the first two are complete, starting everything over because the third is wrong would waste work.&lt;/p&gt;

&lt;p&gt;Be explicit: “Keep the results of the first two items and adjust only the third.” Then provide the actual result, expected result, and relevant location. After the fix, also check whether it affected the completed parts.&lt;/p&gt;

&lt;h2&gt;
  
  
  Use References to Reduce Abstract Descriptions
&lt;/h2&gt;

&lt;p&gt;Suppose you want a hairstylist to cut your bangs. Would you describe the function curve &lt;em&gt;f(x) = x/4 + 1&lt;/em&gt;, or show a reference photo?&lt;/p&gt;

&lt;p&gt;For most people, the latter is easier to communicate. Working with AI is similar: existing screens, components, code examples, and API formats can all serve as concrete references.&lt;/p&gt;

&lt;p&gt;But specify which parts should be reused. For example: “Use only the spacing and button style from this screen as a reference. Follow the current requirements for data fields and interaction flow.” Otherwise, AI may copy unsuitable details from the example as well.&lt;/p&gt;

&lt;h2&gt;
  
  
  Use Screenshots to Locate the Issue, Then Verify Through Actual Behavior
&lt;/h2&gt;

&lt;p&gt;When a project has many similar components and text descriptions keep missing the target, screenshots with marked areas, combined with page paths and reproduction steps, usually make it easier to locate the issue.&lt;/p&gt;

&lt;p&gt;However, a screenshot only shows a screen at one point in time. It cannot, on its own, prove that the data is correct, interactions work, or the problem has been resolved. When necessary, have an agent with appropriate permissions interact with the application and reproduce the issue, then verify the outcome using relevant tests or logs. This reduces the chance of declaring success based on appearance alone.&lt;/p&gt;

&lt;h1&gt;
  
  
  8. Doing Too Much: Let AI Spend Its Effort Where Judgment Is Needed
&lt;/h1&gt;

&lt;p&gt;The situations above have something in common: AI did a lot of work, but some of it did not move the task forward.&lt;/p&gt;

&lt;p&gt;Examples include repeatedly converting the same document, looking up standard commands again, reading entire huge logs, or running unrelated project-wide workflows for a small change.&lt;/p&gt;

&lt;p&gt;The principle is simple: &lt;strong&gt;turn repeatable processes into tools, and use the model for the parts that require judgment.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Turn Repeated Workflows into Tools and Skills
&lt;/h2&gt;

&lt;p&gt;Repeatable work such as file conversion, formatting, data checks, and report generation is well suited to scripts. A Skill can then explain when to use those scripts, what inputs they require, and how to verify the results.&lt;/p&gt;

&lt;p&gt;A Skill is not a zero-cost shortcut. Its value lies in reducing the need to rediscover and orchestrate a workflow every time. The more stable the underlying tools are, the easier it is to accumulate benefits.&lt;/p&gt;

&lt;p&gt;Even before creating a Skill, you can start with project instruction files supported by your tools and an index under &lt;code&gt;docs/&lt;/code&gt;, such as the &lt;code&gt;AGENTS.md&lt;/code&gt; file used by some agents. File names, loading rules, and support vary by tool. Placing a document in a folder does not mean every agent will read it automatically.&lt;/p&gt;

&lt;h2&gt;
  
  
  Multiple Agents: Calculate Handoff Costs Before Assigning Roles
&lt;/h2&gt;

&lt;p&gt;For large tasks, you can borrow from project management and separate requirements clarification, architecture design, implementation, testing, and operational checks.&lt;/p&gt;

&lt;p&gt;But you do not need an agent for every job title. Each agent may still need to read background information, exchange information, and validate results. More roles can mean higher coordination costs. Parallel execution can shorten waiting time without reducing total token usage.&lt;/p&gt;

&lt;p&gt;A more reasonable approach is to divide work according to task boundaries, dependencies, and acceptance criteria. Let a single agent with appropriate tools handle small tasks, and split larger tasks into pieces that can be completed independently.&lt;/p&gt;

&lt;p&gt;The value of agentic engineering lies in clear task decomposition, accurate handoffs, and reliable validation. Whether it is worthwhile should be judged by the total cost and quality of completing the task.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F98vlqvn3wr6sx0umamuc.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F98vlqvn3wr6sx0umamuc.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h1&gt;
  
  
  Conclusion: Saving Tokens Starts with Reducing Rework
&lt;/h1&gt;

&lt;p&gt;Sometimes tokens burn quickly because the task is inherently complex. Other times, it is because we send AI in circles among unclear goals, outdated context, and repetitive work.&lt;/p&gt;

&lt;p&gt;Rather than simply shortening prompts, I care more about clarifying four things before handing over a task: what needs to change, where to start, what constraints apply, and what counts as done.&lt;/p&gt;

&lt;p&gt;Combine that with an appropriate model, concise context, reusable tools, and necessary validation, and each run has a better chance of moving toward the goal.&lt;/p&gt;

&lt;p&gt;What we should measure is the total cost of completing the same task—including tokens, waiting time, and the time spent on human corrections.&lt;/p&gt;

&lt;h1&gt;
  
  
  References
&lt;/h1&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://docs.github.com/en/copilot/how-tos/copilot-in-your-ide/use-copilot-agents/use-plan-mode" rel="noopener noreferrer"&gt;GitHub’s official documentation on Plan mode&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://developers.openai.com/api/docs/guides/prompt-caching" rel="noopener noreferrer"&gt;Prompt caching | OpenAI API&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://platform.claude.com/docs/en/build-with-claude/prompt-caching" rel="noopener noreferrer"&gt;Anthropic’s official pricing information&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;h1&gt;
  
  
  Coming Next: How Do You Give AI a Requirement?
&lt;/h1&gt;

&lt;p&gt;Why do some ideas still require repeated revisions from AI even when we feel we have explained them clearly?&lt;/p&gt;

&lt;p&gt;When people communicate, shared experience often fills in the details left unsaid. Requirements given to AI need those unspoken understandings translated into conditions that can be understood, executed, and verified.&lt;/p&gt;

&lt;p&gt;In the next article, we will explore how to turn the ideas in your head into requirements AI can reliably implement.&lt;/p&gt;




&lt;p&gt;The &lt;a href="https://medium.com/@lian000123/vibe-coding-%E9%96%8B%E7%99%BC%E5%B0%8E%E5%85%A5-ai-%E5%95%8F%E9%A1%8C%E7%AD%86%E8%A8%98-%E4%BA%8C-%E7%82%BA%E4%BB%80%E9%BA%BC-token-%E7%87%92%E5%BE%97%E9%80%99%E9%BA%BC%E5%BF%AB-258a135acd54" rel="noopener noreferrer"&gt;original article&lt;/a&gt; was written in Traditional Chinese, and the Japanese and English versions were translated using ChatGPT.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>softwareengineering</category>
      <category>vibecoding</category>
      <category>tokenoptimization</category>
    </item>
    <item>
      <title>Notes on 3D Modeling Training with GPT-6 AGI</title>
      <dc:creator>Lian0123 (連永立)</dc:creator>
      <pubDate>Sun, 13 Sep 2026 18:45:35 +0000</pubDate>
      <link>https://dev.to/lian0123/notes-on-3d-modeling-training-with-gpt-6-agi-4nlf</link>
      <guid>https://dev.to/lian0123/notes-on-3d-modeling-training-with-gpt-6-agi-4nlf</guid>
      <description>&lt;p&gt;GPT-6 Aatra was recently released, and many people online have been using it to create models with fairly good results. So I decided to give it a try. The result was a little painful to look at, which led me to start experimenting with ways to improve it.&lt;/p&gt;

&lt;h1&gt;
  
  
  The Problem
&lt;/h1&gt;

&lt;p&gt;First, these are the three-view images I fed into the AI, generated by the AI itself:&lt;br&gt;
&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdj46aedkjg9lo6t6sbai.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdj46aedkjg9lo6t6sbai.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;This is the result created by GPT-6 Aatra. It is better than the 5.6 model, but still far from satisfactory:&lt;br&gt;
&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fh9kc8ppiinhap3p7p0k5.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fh9kc8ppiinhap3p7p0k5.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;After trying many different approaches, all of which failed, I suddenly had an idea: could the model’s default understanding of character modeling be fundamentally flawed—almost like a person with a cognitive impairment? If we fed the AI a character-modeling tutorial video, could that remove the obstacle and allow it to create a beautiful model?&lt;/p&gt;

&lt;p&gt;The idea was based on OpenAI’s Video Pre-Training theory. If AI could learn how to perform tasks by watching YouTube videos, like in the Minecraft diamond-mining competition back then, then perhaps videos could help AI fully understand what it was supposed to do.&lt;/p&gt;

&lt;p&gt;For a model with AGI-level capabilities, could we let it watch videos and use what it learned, together with reasoning, to make up for its lack of understanding? And since it is a model approaching AGI, could it learn an entire workflow after watching a video just once?&lt;/p&gt;

&lt;p&gt;As it turns out, it can. Since I only have access to the Plus plan, I could only let it watch two short videos. Even so, it improved the model to the point where it looked like the handiwork of a temporary worker:&lt;br&gt;
&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fvy87zp8zbfsl71qb14ok.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fvy87zp8zbfsl71qb14ok.png" alt=" " width="799" height="327"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Overall, the result was fairly good. The rigging was still incorrect, but considering that I only used Plus and consumed just 60–70% of my weekly token quota, the finished model was already very good. At least, it looked like a person:&lt;br&gt;
&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fmfe4u1xoif8s9xpgoxco.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fmfe4u1xoif8s9xpgoxco.png" alt=" " width="800" height="422"&gt;&lt;/a&gt;&lt;br&gt;
&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fy4ss9x5btrinvi66dkis.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fy4ss9x5btrinvi66dkis.png" alt=" " width="800" height="422"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;I’ve uploaded the files to GitHub. If you’re interested, take a look:&lt;br&gt;&lt;br&gt;
&lt;a href="https://github.com/Lian0123/gpt-6-ai-3d-model-build-demo" rel="noopener noreferrer"&gt;https://github.com/Lian0123/gpt-6-ai-3d-model-build-demo&lt;/a&gt;&lt;/p&gt;

&lt;h1&gt;
  
  
  Optimization
&lt;/h1&gt;

&lt;p&gt;Since GPT-6 can do this, could I distill the process into a skill and make it possible for lower-end models to understand and reproduce it—achieving “token financial freedom”?&lt;/p&gt;

&lt;p&gt;Reality dealt me a heavy blow. The skill was successfully created, and the three-view images were normal at first:&lt;/p&gt;

&lt;p&gt;But the result was still a disaster. Although the body modeling was relatively accurate, the face was still a failure:&lt;/p&gt;

&lt;p&gt;This seems like a very worthwhile topic to investigate. If I get any new results, I’ll update this article.&lt;/p&gt;

&lt;h1&gt;
  
  
  Further Thoughts
&lt;/h1&gt;

&lt;p&gt;If we can rapidly train AGI through videos, could this method also make it possible to handle tasks such as special-effects production, video editing, and even high-end artistic creation?&lt;/p&gt;




&lt;p&gt;The &lt;a href="https://medium.com/@lian000123/%E8%AB%96gpt-6-agi-%E5%BB%BA%E6%A8%A1%E8%A8%93%E7%B7%B4%E7%AD%86%E8%A8%98-f61d50c3c3ab?sharedUserId=lian000123" rel="noopener noreferrer"&gt;original&lt;/a&gt; article was written in Traditional Chinese, and the Japanese and English versions were translated using ChatGPT.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>blennder</category>
      <category>gpt6</category>
      <category>agi</category>
    </item>
    <item>
      <title>Notes on the Pitfalls of iOS Sandbox Accounts and Apple Pay</title>
      <dc:creator>Lian0123 (連永立)</dc:creator>
      <pubDate>Tue, 08 Sep 2026 12:50:39 +0000</pubDate>
      <link>https://dev.to/lian0123/notes-on-the-pitfalls-of-ios-sandbox-accounts-and-apple-pay-61o</link>
      <guid>https://dev.to/lian0123/notes-on-the-pitfalls-of-ios-sandbox-accounts-and-apple-pay-61o</guid>
      <description>&lt;h1&gt;
  
  
  Introduction
&lt;/h1&gt;

&lt;p&gt;Why am I writing this article? The reason is simple: a while ago, a new project at my company required us to integrate Google Pay and Apple Pay.&lt;/p&gt;

&lt;p&gt;The service had already completed its credit card payment flow. However, management later requested support for Google Pay and Apple Pay as well.&lt;/p&gt;

&lt;p&gt;Since none of our previous company projects had integrated Apple Pay, I became the first person in the company to take on this task.&lt;/p&gt;

&lt;p&gt;With the help of AI, the payment integration itself was completed in about one day.&lt;/p&gt;

&lt;p&gt;However, a new problem came up:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;The code was finished, but how were we supposed to test it in the UAT environment?&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;This is basically a record of my experience and the problems I encountered while testing Apple Pay in the Sandbox environment. I hope it helps other engineers who are stuck in a similar testing environment.&lt;/p&gt;

&lt;h1&gt;
  
  
  Problem 1: How to Configure the CER and Private Key
&lt;/h1&gt;

&lt;p&gt;At the moment, I only have a &lt;code&gt;.p12&lt;/code&gt; file. The payment integration side has already configured the Apple Pay certificate and generated the corresponding CSR. The only remaining step is to configure the environment variables on the application side.&lt;/p&gt;

&lt;p&gt;The contents of a &lt;code&gt;.p12&lt;/code&gt; file are generally as follows:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;.p12
├── Certificate
├── Private Key
└── Certificate Chain (may be included)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The overall process is:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;.p12
├── Extract the Certificate / PEM
├── Extract the Private Key
└── Verify that the CSR, Certificate, and Private Key belong to the same key pair
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The following process uses OpenSSL:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# Inspect the p12 contents without outputting the private key&lt;/span&gt;
openssl pkcs12 &lt;span class="nt"&gt;-info&lt;/span&gt; &lt;span class="nt"&gt;-in&lt;/span&gt; apple-pay.p12 &lt;span class="nt"&gt;-noout&lt;/span&gt;

&lt;span class="c"&gt;# Extract the certificate&lt;/span&gt;
openssl pkcs12 &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-in&lt;/span&gt; apple-pay.p12 &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-clcerts&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-nokeys&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  | openssl x509 &lt;span class="nt"&gt;-out&lt;/span&gt; cert.pem

&lt;span class="c"&gt;# Extract the private key&lt;/span&gt;
openssl pkcs12 &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-in&lt;/span&gt; apple-pay.p12 &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-nocerts&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-nodes&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  | openssl pkey &lt;span class="nt"&gt;-out&lt;/span&gt; private-key.pem
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;-nodes&lt;/code&gt; means that the exported private key will not be encrypted again. Therefore, the generated &lt;code&gt;private-key.pem&lt;/code&gt; must be stored securely. It should never be committed to Git or printed in logs.&lt;/p&gt;

&lt;h2&gt;
  
  
  Verifying That the Certificate and Private Key Match
&lt;/h2&gt;

&lt;p&gt;The simplest way is to extract the public key from the certificate and the private key, then compare their hash values.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# Public key from the certificate&lt;/span&gt;
openssl x509 &lt;span class="nt"&gt;-in&lt;/span&gt; cert.pem &lt;span class="nt"&gt;-pubkey&lt;/span&gt; &lt;span class="nt"&gt;-noout&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  | openssl pkey &lt;span class="nt"&gt;-pubin&lt;/span&gt; &lt;span class="nt"&gt;-outform&lt;/span&gt; DER &lt;span class="se"&gt;\&lt;/span&gt;
  | shasum &lt;span class="nt"&gt;-a&lt;/span&gt; 256

&lt;span class="c"&gt;# Public key corresponding to the private key&lt;/span&gt;
openssl pkey &lt;span class="nt"&gt;-in&lt;/span&gt; private-key.pem &lt;span class="nt"&gt;-pubout&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  | openssl pkey &lt;span class="nt"&gt;-pubin&lt;/span&gt; &lt;span class="nt"&gt;-outform&lt;/span&gt; DER &lt;span class="se"&gt;\&lt;/span&gt;
  | shasum &lt;span class="nt"&gt;-a&lt;/span&gt; 256
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If the hash values produced by the two commands are identical, the certificate and private key belong to the same key pair.&lt;/p&gt;

&lt;p&gt;If you also need to verify the CSR, run the following command:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# Public key corresponding to the CSR&lt;/span&gt;
openssl req &lt;span class="nt"&gt;-in&lt;/span&gt; request.csr &lt;span class="nt"&gt;-pubkey&lt;/span&gt; &lt;span class="nt"&gt;-noout&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  | openssl pkey &lt;span class="nt"&gt;-pubin&lt;/span&gt; &lt;span class="nt"&gt;-outform&lt;/span&gt; DER &lt;span class="se"&gt;\&lt;/span&gt;
  | shasum &lt;span class="nt"&gt;-a&lt;/span&gt; 256
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If the hash values of the CSR, certificate, and Private Key are all identical, it means that all three use the same key pair.&lt;/p&gt;

&lt;h2&gt;
  
  
  Architecture
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fegcadqd5tu1i04jwodkq.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fegcadqd5tu1i04jwodkq.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Terminology
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;p12&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;code&gt;.p12&lt;/code&gt; is a PKCS#12 container file. It can usually contain:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Certificate&lt;/li&gt;
&lt;li&gt;Private Key&lt;/li&gt;
&lt;li&gt;Certificate Chain&lt;/li&gt;
&lt;li&gt;Password protecting the private key&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Therefore, under normal circumstances, the &lt;code&gt;.p12&lt;/code&gt; file already contains the certificate and private key required by the payment service.&lt;/p&gt;

&lt;p&gt;However, you still need to confirm the purpose of the &lt;code&gt;.p12&lt;/code&gt; file. Common Apple Pay certificates include:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;Merchant Identity Certificate&lt;/code&gt;: Used to authenticate Merchant Sessions for Apple Pay on the Web.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;Payment Processing Certificate&lt;/code&gt;: Used to encrypt payment data. Decryption is usually handled by the merchant or the PSP (Payment Service Provider).&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;These two certificates have different purposes. You cannot assume that every &lt;code&gt;.p12&lt;/code&gt; file can be used as every type of Apple Pay certificate.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;PEM&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;PEM stands for Privacy-Enhanced Mail. It is a text-based encoding format. Certificates, private keys, and CSRs may all be stored in PEM format.&lt;/p&gt;

&lt;p&gt;Example of a certificate PEM:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;-----BEGIN CERTIFICATE-----
...
...
...
-----END CERTIFICATE-----
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Private Key&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;A private key is also usually stored in PEM format.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;-----BEGIN PRIVATE KEY-----
...
...
...
-----END PRIVATE KEY-----
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;CSR&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;CSR stands for Certificate Signing Request.&lt;/p&gt;

&lt;p&gt;The usual process is to generate a CSR first, then upload it to Apple Developer to obtain a CER.&lt;/p&gt;

&lt;p&gt;A CSR contains the public key, application information, and a signature generated using the corresponding Private Key. However, it does not contain the Private Key itself.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;-----BEGIN CERTIFICATE REQUEST-----
...
...
...
-----END CERTIFICATE REQUEST-----
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;CER&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;A CER is a certificate file issued by Apple. Its file extension is usually &lt;code&gt;.cer&lt;/code&gt;.&lt;/p&gt;

&lt;h1&gt;
  
  
  Problem 2: The Apple Pay Screen Works, but the Card Cannot Be Added
&lt;/h1&gt;

&lt;h2&gt;
  
  
  Situation
&lt;/h2&gt;

&lt;p&gt;The Apple Pay configuration for the website was complete, and the Apple Pay button was displayed correctly. Clicking the button also opened the Apple Pay payment screen.&lt;/p&gt;

&lt;p&gt;I then opened the &lt;a href="https://developer.apple.com/apple-pay/sandbox-testing/" rel="noopener noreferrer"&gt;Apple Pay Sandbox Testing page&lt;/a&gt; and tried to add Visa, Mastercard, and JCB test cards to Wallet.&lt;/p&gt;

&lt;p&gt;Many iOS testing guides mention enabling Developer Mode and adding the newly created Sandbox Account to the Sandbox settings. However, none of the cards could be added.&lt;/p&gt;

&lt;p&gt;Many guides introduce the following steps:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Enable Developer Mode.&lt;/li&gt;
&lt;li&gt;Create a Sandbox Account.&lt;/li&gt;
&lt;li&gt;Add the Sandbox Account to the test settings.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;In reality, these steps are not enough.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Is Developer Mode?
&lt;/h2&gt;

&lt;p&gt;Developer Mode allows the device to run apps that are still under development.&lt;/p&gt;

&lt;p&gt;Enabling Developer Mode does not automatically switch the device to Apple Pay Sandbox mode, nor does it mean that test cards can be added directly to Wallet.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Is a Sandbox Account?
&lt;/h2&gt;

&lt;p&gt;A Sandbox Account is a test account created in App Store Connect. It is used to test Apple Pay, In-App Purchases, and other sandbox features.&lt;/p&gt;

&lt;p&gt;It can usually be created here:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;App Store Connect
→ Users and Access
→ Sandbox
→ Testers
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;When creating the account, make sure that the Email address has not already been registered as a regular Apple Account. Using an existing Email address may cause login or verification problems.&lt;a href="https://developer.apple.com/help/app-store-connect/test-in-app-purchases/create-a-sandbox-apple-account/" rel="noopener noreferrer"&gt;Official Apple Sandbox Account documentation&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The Correct Apple Pay Testing Process
&lt;/h2&gt;

&lt;p&gt;The general Apple Pay Sandbox testing process is as follows:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Create a Sandbox Tester Account in App Store Connect.&lt;/li&gt;
&lt;li&gt;Confirm that the test device supports Apple Pay.&lt;/li&gt;
&lt;li&gt;Set the device region to an Apple Pay-supported region, such as Taiwan.&lt;/li&gt;
&lt;li&gt;Sign out of the Apple Account or iCloud account previously used on the test device.&lt;/li&gt;
&lt;li&gt;Sign in to the test device using the Sandbox Account.&lt;/li&gt;
&lt;li&gt;Open Wallet.&lt;/li&gt;
&lt;li&gt;Select the option to add a credit or debit card.&lt;/li&gt;
&lt;li&gt;Manually enter the test card information provided on Apple’s official page.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Apple’s test card page mainly provides test FPANs, CVVs, CVCs, and expiration dates. It is not a page that directly adds cards to Wallet.&lt;/p&gt;

&lt;p&gt;It is also recommended to use a physical device for Apple Pay testing. Make sure the device has a passcode, Face ID, or Touch ID configured.&lt;/p&gt;

&lt;p&gt;Apple’s official process is to sign out of the existing iCloud account, sign in to the device using the Sandbox Tester Account, and then manually add the test card from Wallet.&lt;a href="https://developer.apple.com/apple-pay/sandbox-testing/" rel="noopener noreferrer"&gt;Apple Pay Sandbox Testing&lt;/a&gt;&lt;/p&gt;

&lt;h1&gt;
  
  
  Problem 3: iCloud 34608 Appears After Signing In to the Sandbox Account
&lt;/h1&gt;

&lt;h2&gt;
  
  
  Situation
&lt;/h2&gt;

&lt;p&gt;Some people may have already signed out of their original Apple Account and attempted to sign in with a Sandbox Account.&lt;/p&gt;

&lt;p&gt;However, after entering the account and password, the iPhone may remain stuck on the account setup animation and eventually display the iCloud 34608 error.&lt;/p&gt;

&lt;p&gt;This error message is not very informative. Apple has not published a complete official explanation for error 34608, and the information available online is scattered. Therefore, the following is the approach I used to troubleshoot the problem.&lt;/p&gt;

&lt;p&gt;First, make sure that the registered Email address is new and has never been registered as a regular Apple Account. If you have made it this far, I believe most readers are probably not stuck at this step.&lt;/p&gt;

&lt;p&gt;After confirming that the Email address is new, log in to the &lt;a href="https://account.apple.com/" rel="noopener noreferrer"&gt;Apple Account&lt;/a&gt; management page using a browser.&lt;/p&gt;

&lt;p&gt;If the system asks you to complete any of the following verification steps, make sure to complete them:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Phone number verification&lt;/li&gt;
&lt;li&gt;Two-factor authentication&lt;/li&gt;
&lt;li&gt;Account security settings&lt;/li&gt;
&lt;li&gt;Acceptance of the terms and conditions&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This is a common problem in corporate testing environments. A company may use a test device that has not completed phone verification or multi-factor authentication. This can cause an iCloud error during account login and prevent the Apple Account from being added to the iPhone.&lt;/p&gt;

&lt;p&gt;After completing the verification steps above, return to the test device and sign in to the Sandbox Account again. If it still fails, restart the device, confirm that the original Apple Account has been signed out, and try again.&lt;/p&gt;

&lt;h1&gt;
  
  
  Final Testing Checklist
&lt;/h1&gt;

&lt;p&gt;If you are testing Apple Pay Sandbox from scratch, I recommend checking the following items in order:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The Merchant ID has been created.&lt;/li&gt;
&lt;li&gt;The purpose of the Apple Pay certificate has been confirmed.&lt;/li&gt;
&lt;li&gt;The CSR, certificate, and Private Key have been verified to belong to the same key pair.&lt;/li&gt;
&lt;li&gt;A Sandbox Tester Account has been created in App Store Connect.&lt;/li&gt;
&lt;li&gt;The Email address used by the Sandbox Account has never been registered as a regular Apple Account.&lt;/li&gt;
&lt;li&gt;The test device supports Apple Pay.&lt;/li&gt;
&lt;li&gt;Developer Mode is enabled on the test device.&lt;/li&gt;
&lt;li&gt;The test device region is set to an Apple Pay-supported region.&lt;/li&gt;
&lt;li&gt;The original Apple Account or iCloud account has been signed out.&lt;/li&gt;
&lt;li&gt;The test device is signed in with the Sandbox Account.&lt;/li&gt;
&lt;li&gt;The official Apple test card has been manually added to Wallet.&lt;/li&gt;
&lt;li&gt;The Apple Pay Web Merchant Domain has been verified.&lt;/li&gt;
&lt;li&gt;The website uses HTTPS and the server’s TLS configuration has been checked.&lt;/li&gt;
&lt;li&gt;Sandbox and Production payment environments are not mixed.&lt;/li&gt;
&lt;li&gt;Payment tokens and Private Keys are not printed in logs.&lt;/li&gt;
&lt;/ul&gt;

&lt;h1&gt;
  
  
  Conclusion
&lt;/h1&gt;

&lt;p&gt;This article mainly covered the following three points:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;A &lt;code&gt;.p12&lt;/code&gt; file is a container, not simply a certificate. You need to extract the certificate and Private Key from it and verify that they belong to the same key pair.&lt;/li&gt;
&lt;li&gt;Apple Pay Sandbox is not just about enabling Developer Mode. You also need to create a Sandbox Account correctly, sign in to the test device, and manually add a test card through Wallet.&lt;/li&gt;
&lt;li&gt;If iCloud 34608 appears, the problem may not be in the Apple Pay code. It may be caused by incomplete Sandbox Account verification or an incomplete device login process.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The most misleading part of Apple Pay testing is assuming that “the payment screen appears” means the entire payment flow is working correctly.&lt;/p&gt;

&lt;p&gt;Certificates, accounts, Wallet, Apple Pay Sessions, and the final decryption of payment data all have their own testing requirements.&lt;/p&gt;

&lt;p&gt;By checking each stage separately, you can usually identify the actual source of the problem much faster.&lt;/p&gt;




&lt;p&gt;The &lt;a href="https://medium.com/@lian000123/ios-sandbox-account-%E8%88%87-apple-pay%E7%9A%84%E8%B8%A9%E5%9D%91%E7%AD%86%E8%A8%98-%E5%90%AB-icloud-error-34608%E5%95%8F%E9%A1%8C-f7de82aa3a40" rel="noopener noreferrer"&gt;original article&lt;/a&gt; was written in Traditional Chinese, and the Japanese and English versions were translated using ChatGPT.&lt;/p&gt;

</description>
      <category>applepay</category>
      <category>sandbox</category>
      <category>testing</category>
      <category>icloud34608</category>
    </item>
    <item>
      <title>Notes on Introducing AI into Vibe Coding Development (Part 1): Why Does AI “Cut Corners”?</title>
      <dc:creator>Lian0123 (連永立)</dc:creator>
      <pubDate>Sun, 19 Jul 2026 04:03:02 +0000</pubDate>
      <link>https://dev.to/lian0123/notes-on-introducing-ai-into-vibe-coding-part-1-why-does-ai-cut-corners-3kma</link>
      <guid>https://dev.to/lian0123/notes-on-introducing-ai-into-vibe-coding-part-1-why-does-ai-cut-corners-3kma</guid>
      <description>&lt;h2&gt;
  
  
  Introduction
&lt;/h2&gt;

&lt;p&gt;From Vibe Coding to Harness Engineering, integrating AI into software development has become one of the industry’s most discussed topics. Many companies are adopting AI to improve productivity, while new models and tools optimized for programming continue to emerge.&lt;/p&gt;

&lt;p&gt;Choosing the right model can save you from many unnecessary detours. But if your prompt is poorly designed, AI may lead you even further away from your actual goal.&lt;/p&gt;

&lt;p&gt;This article summarizes several common mistakes I have observed while introducing AI into development workflows. These principles apply not only to everyday prompting, but also to the design of reusable skills, project rules, and engineering workflows.&lt;/p&gt;

&lt;p&gt;Keep one thing in mind: when AI starts to “cut corners,” humans usually pay for it with more tokens and more time spent fixing the result.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F001qb3mmzvgxujfknvah.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F001qb3mmzvgxujfknvah.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does AI “Cut Corners”?
&lt;/h2&gt;

&lt;p&gt;Anyone familiar with deep learning may have heard of &lt;strong&gt;specification gaming&lt;/strong&gt; or &lt;strong&gt;reward hacking&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;In simple terms, an AI system may exploit gaps in an instruction or evaluation rule to satisfy the visible objective at the lowest possible cost. The task may appear complete, while the output is useless—or even contradicts the real intention behind the request.&lt;/p&gt;

&lt;h3&gt;
  
  
  A Simple Example
&lt;/h3&gt;

&lt;p&gt;Imagine a student named Xiaoming who receives 100 units of allowance every day. To encourage him to study, his parents require him to take a practice exam each day:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;If his score drops by 10 points, his allowance is reduced by 10 units.&lt;/li&gt;
&lt;li&gt;If his score rises by 10 points, his allowance increases by 10 units.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Xiaoming scores zero on the first day, so his allowance drops to 90 units. He soon realizes, however, that if he continues scoring zero every day, his score will never fall below the previous day’s result. He can keep receiving the full 100 units without studying.&lt;/p&gt;

&lt;p&gt;Xiaoming has not violated the rules, but the system has completely failed to achieve its purpose. His parents wanted him to study, yet they accidentally created an optimal strategy in which he can avoid studying and still receive a stable allowance.&lt;/p&gt;

&lt;h3&gt;
  
  
  A Software Engineering Example
&lt;/h3&gt;

&lt;p&gt;Suppose an engineer asks an AI agent to “fix the bug in this program.”&lt;/p&gt;

&lt;p&gt;Without additional constraints, the AI could theoretically delete the faulty code—or remove the entire feature. If the program no longer exists, the error can no longer occur. That technically eliminates the bug, but it clearly does not satisfy the user’s real need.&lt;/p&gt;

&lt;p&gt;In Vibe Coding, AI often gravitates toward the fastest, cheapest, and easiest-to-mark-as-complete solution. If a prompt describes only the surface-level objective without defining the intent, constraints, and acceptance criteria, the result can easily diverge from expectations.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F07smdjayeljfy57x9fc8.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F07smdjayeljfy57x9fc8.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Improving AI Behavior Through Better Instructions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Method 1: Provide a Reference Example
&lt;/h3&gt;

&lt;p&gt;One of the most effective approaches is to provide a correct example and ask the AI to follow an established pattern.&lt;/p&gt;

&lt;p&gt;The goal is not to prevent the AI from reasoning. It is to reduce unnecessary guesswork. When naming conventions, file structure, interface style, and implementation patterns are demonstrated clearly, the AI is less likely to misunderstand the task, make implementation mistakes, or hallucinate nonexistent behavior.&lt;/p&gt;

&lt;p&gt;This approach is especially useful for:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;interface development in large projects;&lt;/li&gt;
&lt;li&gt;building a new page based on an existing one;&lt;/li&gt;
&lt;li&gt;integrating third-party APIs;&lt;/li&gt;
&lt;li&gt;reusing established components, tests, or data formats;&lt;/li&gt;
&lt;li&gt;implementing features with complete official documentation.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Instead of saying, “Build me a new page,” say, “Implement page B by following page A’s directory structure, component boundaries, and error-handling pattern.”&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2wjvu39910kerxz3cdsw.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2wjvu39910kerxz3cdsw.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  Method 2: State Explicit Prohibitions
&lt;/h3&gt;

&lt;p&gt;Another common technique is to turn known risks and failure modes into explicit prohibitions.&lt;/p&gt;

&lt;p&gt;For example:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Do not remove existing functionality.&lt;/li&gt;
&lt;li&gt;Do not disable tests to make the test suite pass.&lt;/li&gt;
&lt;li&gt;Do not replace a production API with mock data unless clearly instructed.&lt;/li&gt;
&lt;li&gt;Do not change the public interface.&lt;/li&gt;
&lt;li&gt;Do not place keys, tokens, or personal data in source code.&lt;/li&gt;
&lt;li&gt;Do not perform a large refactor before confirming the impact.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This reflects a useful practice in AI-assisted development: &lt;strong&gt;turn previous mistakes into reusable rules&lt;/strong&gt;. Whether those rules live in &lt;code&gt;AGENTS.md&lt;/code&gt;, &lt;code&gt;CLAUDE.md&lt;/code&gt;, a skill, or a team knowledge base, the purpose is the same—to prevent known failures from happening again.&lt;/p&gt;

&lt;p&gt;This method works well when the risks and operational boundaries are already understood. However, more prohibitions do not automatically produce better results. Too many rules, conflicting rules, or rules without priorities can overload the context and bury the constraints that matter most.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F6w034y32ntlfjx57ieai.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F6w034y32ntlfjx57ieai.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  Method 3: Require a Detailed Plan First
&lt;/h3&gt;

&lt;p&gt;One of the central benefits of plan modes and skills is that they break work into reviewable steps.&lt;/p&gt;

&lt;p&gt;A useful implementation plan should include at least:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;goals and non-goals;&lt;/li&gt;
&lt;li&gt;the expected scope of changes;&lt;/li&gt;
&lt;li&gt;implementation order;&lt;/li&gt;
&lt;li&gt;potential risks;&lt;/li&gt;
&lt;li&gt;testing and verification methods;&lt;/li&gt;
&lt;li&gt;completion criteria.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;When the workflow is written down, humans can identify misunderstandings before the AI begins making changes. This also reduces the chance of unexpected actions during execution.&lt;/p&gt;

&lt;p&gt;This approach is particularly suitable for new projects, large refactors, and feature development in projects with an established knowledge base.&lt;/p&gt;

&lt;p&gt;Its limitation is equally clear: the quality of the plan depends on the completeness of the available information. If an important condition is missing during planning, the AI may faithfully execute a plan that looks complete but is fundamentally wrong.&lt;/p&gt;

&lt;p&gt;A more mature approach is to maintain the following information at the documentation level:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;development rules;&lt;/li&gt;
&lt;li&gt;architecture and feature responsibilities;&lt;/li&gt;
&lt;li&gt;change history;&lt;/li&gt;
&lt;li&gt;risk assessments;&lt;/li&gt;
&lt;li&gt;acceptance criteria;&lt;/li&gt;
&lt;li&gt;known issues and the reasoning behind past decisions.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Together, these materials form a living project knowledge base. I will explore this topic in a future article on &lt;strong&gt;AI Living Documentation&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F253455kj9m07eb2nnxmj.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F253455kj9m07eb2nnxmj.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  Method 4: Define the Expected Outcome Clearly
&lt;/h3&gt;

&lt;p&gt;If you can describe exactly what “done” should look like, the AI’s output is more likely to converge on the intended result.&lt;/p&gt;

&lt;p&gt;For front-end work, this may include a wireframe, design mockup, interaction flow, target screenshot, or existing page. For back-end work, it may include an API specification, input and output examples, error codes, and performance requirements.&lt;/p&gt;

&lt;p&gt;In addition to explaining what must be built, define:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;which existing behaviors must remain unchanged;&lt;/li&gt;
&lt;li&gt;which scenarios must succeed;&lt;/li&gt;
&lt;li&gt;which errors must be handled correctly;&lt;/li&gt;
&lt;li&gt;how completion will be demonstrated.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Clear acceptance criteria prevent the AI from treating “the code has been written” as equivalent to “the feature is complete.”&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fcmw9mu9hb84wngmugrbe.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fcmw9mu9hb84wngmugrbe.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Common Problems in Practice
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Problem 1: Claiming Completion Without Verification
&lt;/h3&gt;

&lt;p&gt;Suppose a user asks an AI agent to integrate a particular library. If the model misunderstands the library’s version, API, or documentation, the generated code may look plausible while failing at runtime.&lt;/p&gt;

&lt;p&gt;The original instruction might be:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Add Live2D to this web page so that the character on the page can be animated.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;This describes the feature but does not require the AI to verify the result. A stronger instruction would be:&lt;/p&gt;

&lt;blockquote&gt;
&lt;ol&gt;
&lt;li&gt;Add Live2D to this web page so that the character on the page can be animated.
&lt;/li&gt;
&lt;li&gt;Confirm that the Live2D assets load correctly, the model is visible, and its animation plays.
&lt;/li&gt;
&lt;li&gt;Run the project and verify that no related errors appear in the browser console.
&lt;/li&gt;
&lt;li&gt;Do not break existing page functionality. If any item cannot be verified, clearly identify it and explain why.
&lt;/li&gt;
&lt;li&gt;When finished, list the changed files, verification method, and test results.&lt;/li&gt;
&lt;/ol&gt;
&lt;/blockquote&gt;

&lt;p&gt;“Please ensure code quality” is still too abstract. A better approach is to translate quality into executable and observable conditions:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;the project builds successfully;&lt;/li&gt;
&lt;li&gt;automated tests pass;&lt;/li&gt;
&lt;li&gt;the feature works in the specified environment;&lt;/li&gt;
&lt;li&gt;no new errors appear in the browser console;&lt;/li&gt;
&lt;li&gt;existing functionality remains intact;&lt;/li&gt;
&lt;li&gt;any unperformed verification is clearly disclosed.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The key principle is that &lt;strong&gt;development, testing, and verification are separate steps performed in sequence&lt;/strong&gt;. Writing the code does not mean the tests have passed, and passing tests does not necessarily mean the user’s need has been satisfied.&lt;/p&gt;

&lt;p&gt;If this workflow is used repeatedly, encode it in a skill, project rule, or standard checklist instead of relying on an ad hoc reminder every time.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftwc0mebs5xev9ixrvvg3.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftwc0mebs5xev9ixrvvg3.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  Problem 2: Ambiguous Terms Create Divergent Interpretations
&lt;/h3&gt;

&lt;p&gt;Humans cannot read minds, and neither can AI.&lt;/p&gt;

&lt;p&gt;When a requirement is unclear, the AI can only fill in the missing information using its prior knowledge, the current context, and common conventions. Its answer may be reasonable without being what the user actually wanted.&lt;/p&gt;

&lt;p&gt;For example, “menu” in an application could refer to:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;the main menu;&lt;/li&gt;
&lt;li&gt;the settings menu;&lt;/li&gt;
&lt;li&gt;a context menu;&lt;/li&gt;
&lt;li&gt;the top navigation bar;&lt;/li&gt;
&lt;li&gt;a mobile side menu.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Consider this instruction:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Add an “Open File” option to the menu.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The AI has to infer which menu should be changed. Even if it chooses the wrong location, it may still conclude that the task is complete.&lt;/p&gt;

&lt;p&gt;A clearer version would be:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;In the desktop application’s top-level File menu, add an “Open File” option directly below “New File.” When clicked, it should open the system file picker and allow the user to select a &lt;code&gt;.json&lt;/code&gt; file. Cancelling the picker must not display an error. Do not modify the context menu or the settings page.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The more concrete the description, the fewer semantic branches remain—and the less likely the AI is to make the wrong assumption.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftq0xnrc9m1aq8ubgwd43.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftq0xnrc9m1aq8ubgwd43.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Conclusion
&lt;/h2&gt;

&lt;p&gt;When AI “cuts corners,” the cause is often not attitude but incomplete definitions of the goal, constraints, and acceptance criteria.&lt;/p&gt;

&lt;p&gt;If you tell an AI only what to finish, it may choose the shortest path. If you also explain why the task matters, what must not be broken, how the result should be verified, and what counts as complete, the quality of its output becomes far more consistent.&lt;/p&gt;

&lt;p&gt;An effective AI development workflow is not simply a longer prompt. It is a system that continuously accumulates reference examples, prohibitions, implementation plans, acceptance criteria, testing procedures, and project knowledge.&lt;/p&gt;

&lt;h2&gt;
  
  
  Next: Why Am I Burning Through Tokens So Quickly?
&lt;/h2&gt;

&lt;p&gt;Have you ever asked AI to solve one simple problem, only to watch the token usage grow until your wallet starts to worry? The reason may not be that the model talks too much. Your development process may be repeatedly paying the cost of misunderstandings.&lt;/p&gt;

&lt;p&gt;Stay tuned for the next article.&lt;/p&gt;




&lt;p&gt;The &lt;a href="https://medium.com/@lian000123/vibe-coding-%E9%96%8B%E7%99%BC%E5%B0%8E%E5%85%A5ai%E5%95%8F%E9%A1%8C%E7%AD%86%E8%A8%98-%E4%B8%80-%E7%82%BA%E4%BB%80%E9%BA%BCai%E6%9C%83%E5%81%B7%E6%87%B6-934a70def12f?sharedUserId=lian000123" rel="noopener noreferrer"&gt;original article&lt;/a&gt; was written in Traditional Chinese, and the Japanese and English versions were translated using ChatGPT.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>vibecoding</category>
      <category>softwareengineering</category>
      <category>development</category>
    </item>
    <item>
      <title>Notes on Introducing AI into Vibe Coding Development (Part 0) - Why I Am Writing This Article</title>
      <dc:creator>Lian0123 (連永立)</dc:creator>
      <pubDate>Thu, 02 Jul 2026 16:57:35 +0000</pubDate>
      <link>https://dev.to/lian0123/notes-on-introducing-ai-into-vibe-coding-development-part-zero-why-i-am-writing-this-article-gei</link>
      <guid>https://dev.to/lian0123/notes-on-introducing-ai-into-vibe-coding-development-part-zero-why-i-am-writing-this-article-gei</guid>
      <description>&lt;h2&gt;
  
  
  What Is Happening in the Software Industry Now?
&lt;/h2&gt;

&lt;p&gt;In recent years, with the emergence of general-purpose artificial intelligence such as ChatGPT and Claude Code, often referred to as AGI, companies have increasingly begun introducing AI into their workflows.&lt;/p&gt;

&lt;p&gt;Starting in June 2026, GitHub Copilot AI made a major change. It moved away from the PRU model, or premium request units, which was based on request counts, and shifted toward a model calculated using AI Credits. As a result, the execution cost for an engineer developing a feature, reviewing PRs, organizing documentation, and performing other tasks has gradually increased.&lt;/p&gt;

&lt;p&gt;Some developers even used up an entire month’s allowance on the very first day GitHub Copilot AI adopted the AI Credit model. This shift toward token-based pricing is becoming increasingly mainstream. It also places more restrictions on using AI within workflows. If you are not careful, your usage can run out quickly, and costs can continue to rise. Worse still, this type of pricing model is likely to become a broader trend among AI service providers.&lt;/p&gt;

&lt;p&gt;Jensen Huang once said:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The question today is no longer whether we should use AI. After all, not every company is like NVIDIA, with the freedom to consume tokens at scale. The real issue is the conflict between problem-solving, money, and time.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fxksuwqnr6ldlwjiao6bu.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fxksuwqnr6ldlwjiao6bu.png" alt="Problem-solving and cost have already become opposing forces." width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Solving problems and controlling costs have already become conflicting forces. The future direction of the software market will be about solving problems effectively while keeping costs under control. It will not be about how many tokens you used to solve something, but about achieving the greatest benefit with the lowest possible cost.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Should Business Managers and Development Leads Do Next?
&lt;/h2&gt;

&lt;p&gt;Clearly, the first step is to identify what qualities and capabilities future technical talent should have:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The ability to break down large problems and requirements, extract the key points, group them properly, and find the best solution&lt;/li&gt;
&lt;li&gt;The ability to simplify AI behavior so it can be executed at the lowest possible cost&lt;/li&gt;
&lt;li&gt;The ability to achieve maximum benefit with the fewest tokens&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Here is an evaluation framework for companies, including both positive and negative evaluation points.&lt;/p&gt;

&lt;h3&gt;
  
  
  Positive Evaluation Points
&lt;/h3&gt;

&lt;p&gt;If the company is willing to spend money to solve problems:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Has this problem been troubling the company for a long time? Was it difficult to execute through conventional methods because of staffing, scheduling, or other reasons?&lt;/li&gt;
&lt;li&gt;Has AI effectively replaced existing talent resources in the way an expert system might? While ensuring correctness, has it saved human effort?&lt;/li&gt;
&lt;li&gt;Was the problem solved quickly? Did it create additional time costs?&lt;/li&gt;
&lt;/ol&gt;

&lt;h3&gt;
  
  
  Negative Evaluation Points
&lt;/h3&gt;

&lt;p&gt;If the company is relatively conservative:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Will this adjustment reduce future changes, or will it increase them? How should we limit the damage?&lt;/li&gt;
&lt;li&gt;If problems occur, are the losses acceptable? For example: time, reputation, and financial cost.&lt;/li&gt;
&lt;li&gt;How much benefit can this adjustment bring? Compared with problem-solving costs and personnel costs, how much money does it actually save?&lt;/li&gt;
&lt;li&gt;Will future maintenance costs decrease? Or will AI make them higher and unnecessarily so?&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;In practice, what business managers need even more is a cost plan:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;How much cost are we willing to invest in solving the problem?&lt;/li&gt;
&lt;li&gt;Is the risk of cost increase controllable?&lt;/li&gt;
&lt;li&gt;Are there audit points in place to control costs?&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Regarding the common online discussion that AI will replace software engineers, my view is relatively optimistic. There are many detailed technical issues that bosses or non-engineering project managers simply cannot fully grasp. When a company truly lays off all of its software talent, that is when the real disaster begins.&lt;/p&gt;

&lt;p&gt;That said, it is undeniable that Junior Engineers will become increasingly rare in the future. The technical threshold will become higher, because using AI well requires the corresponding knowledge. In contrast, Senior Engineers who can combine experience and expertise will become one of the strongest resources in the market.&lt;/p&gt;

&lt;p&gt;The knowledge and experience of software professionals will become an important dividing line.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fgjdzp1kmora2wzyrg54f.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fgjdzp1kmora2wzyrg54f.png" alt="The knowledge and experience of software professionals will become an important dividing line." width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Here is a simple example. Suppose a project developed through Vibe Coding suddenly encounters a serious bug in production. Every minute the problem exists causes the company to lose money.&lt;/p&gt;

&lt;p&gt;If you were the manager, would you want the developer to search for the answer as if rolling dice? Or would you want the developer to quickly identify the root cause?&lt;/p&gt;

&lt;p&gt;I also like this saying:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;People can never imagine what lies outside their own understanding.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The knowledge of a project member is the key to future product maintenance.&lt;/p&gt;

&lt;p&gt;Felix Rieseberg, Engineering Lead at Claude Cowork, also said:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;The biggest barrier to AI adoption is people not realizing they can ask AI to solve almost any problem… The gap isn't technical; it's psychological.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Felix Rieseberg points out that the reason people fail to use AI well lies in imagination. What you can think of determines what you can ask. When using today’s top-tier models, the core problem is not always that the result falls short of expectations. More often, it is that the user has not described the requirements clearly enough, causing the AI to produce results that do not match what was expected.&lt;/p&gt;

&lt;p&gt;Of course, the input and output limitations of model context, as well as the model’s byte-processing issues, must first be addressed.&lt;/p&gt;

&lt;h2&gt;
  
  
  Next Article: Why Does AI Get Lazy?
&lt;/h2&gt;

&lt;p&gt;The next article will discuss a common phenomenon:&lt;/p&gt;

&lt;p&gt;Why is it that when two teams use the same AI, one team significantly improves its efficiency, while the other sees almost no improvement in productivity?&lt;/p&gt;

&lt;p&gt;Many times, it is not that AI does not understand. Rather, it is...&lt;/p&gt;

&lt;p&gt;Stay tuned.&lt;/p&gt;




&lt;p&gt;The &lt;a href="https://medium.com/@lian000123/vibe-coding-%E9%96%8B%E7%99%BC%E5%B0%8E%E5%85%A5ai%E5%95%8F%E9%A1%8C%E7%AD%86%E8%A8%98-%E9%9B%B6-%E7%82%BA%E4%BB%80%E9%BA%BC%E8%A6%81%E5%AF%AB%E9%80%99%E7%AF%87%E6%96%87%E7%AB%A0-1b48d516f61d" rel="noopener noreferrer"&gt;original article&lt;/a&gt; was written in Traditional Chinese, and the Japanese and English versions were translated using ChatGPT.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>softwareengineering</category>
      <category>vibecoding</category>
      <category>development</category>
    </item>
  </channel>
</rss>
