A Brief Discussion on Mainstream AI Service Ecosystems and Related Tools

TLDR

• Developer Ecosystem (API): Data is not used for training by default; suitable for automated integration.
• Consumer Ecosystem (Web): Data may be used for training by default; must be manually disabled.
• Claude: Excels at coding and long-form logic, but conversation history consumes tokens; be aware of the "Claude Code and Web version shared quota" trap.
• Google AI: The ecosystem is complex; the Gemini Advanced and Google One subscription mechanisms are often confusing, and quotas are calculated separately across different platforms (App/NotebookLM/CLI), making it suitable for flexible use across multiple carriers.
• NotebookLM Presentation Limits: 15-page limit per generation; it is recommended to generate in sections and merge them later. Watermarks and image-based text can be handled via Canva or DeckEdit.
• Model Attention Limits: Models tend to miss details from the beginning of long conversations; it is recommended to explicitly restate key information and limit the scope of modifications.
• Agent Editor Selection: Antigravity is suitable for back-and-forth discussions, GitHub Copilot for implementation tasks, and Gemini CLI for simple, repetitive tasks.
• WSL Environment Limitations: Copilot CLI may face issues when executing PowerShell scripts in a WSL environment; using the VS Code extension is recommended.

Mainstream AI Service Architecture

AI services are generally divided into three categories, each with different billing and privacy policies:

• Developer Ecosystem: Primarily based on API Keys, billed by usage; data is not used for training by default (except for the Google AI Studio free tier).
• Enterprise and Team Ecosystem: Provides unified management and SLA guarantees; contracts explicitly prohibit data training.
• Consumer Ecosystem: Divided into free and subscription tiers; Web conversation data may be used for training by default and must be manually disabled in settings.

Anthropic Ecosystem and Claude Code

When you might encounter issues: During frequent local development using Claude Code.

• Quota Calculation Trap: Claude counts the "entire conversation history" into the Context Window when replying; the longer the conversation, the faster tokens are consumed.
• Shared Quota Risk: Claude Code and the web version share the same quota; heavy testing may render the web version unusable.
• Artifacts Advantage: Solves issues with code block truncation in traditional chat interfaces, messy copy-paste formatting, and the inclusion of AI filler text.

Google Ecosystem and NotebookLM

When you might encounter issues: When summarizing long documents or generating presentations.

• NotebookLM Presentation Limits: 15-page limit per generation.
  • Solution: Generate an outline first, create data sources for each section of the transcript, generate them separately, and then merge.
• Watermarks and Image-based Text:
  • Watermark removal: Use Canva's "Magic Eraser" or a dedicated tool like NotebookLM Watermark Remover.
  • Text editing: Use DeckEdit to restore image-based presentations into editable PPTX files.
• Family Sharing Mechanism: Google One's "Cloud Storage" and "AI Credits" are shared by the whole family; individual member usage cannot be restricted.

Model Usage Experience and Common Anomalies

Observations on Model Personality

• ChatGPT: Prone to rambling; it is recommended to use system instructions to lower its enthusiasm.
• Claude: Overly neutral; prone to changing its stance based on new information; it is recommended to provide sufficient context.
• Gemini: High desire to perform; often overly confident when fixing bugs, even when the execution fails.

File Uploads and Web Search Pitfalls

When you might encounter issues: When uploading large files to the web version or relying on AI web searches.

• File Parsing Truncation: The Gemini web version may truncate long Markdown files; it is recommended to paste the text content directly.
• Web Search Mechanism: The model searches via a search engine; if a site is not indexed (e.g., a newly created GitHub Pages site), the model will be unable to read it. It is recommended to submit a sitemap.xml via Google Search Console.
• Gemini Search Logic: If reading fails, Gemini sometimes tends to fabricate content rather than admitting it cannot read the source; pay special attention to this.

AI Agent Editor Practice

Tool Positioning and Selection

• Antigravity: Suitable for tasks requiring "back-and-forth discussion," but be mindful of Claude model quota limits.
• GitHub Copilot: Suitable for assigned "implementation tasks," with costs controlled through iteration limits.
• Gemini CLI: Suitable for simple, high-frequency, repetitive background tasks.

WARNING

Gemini Model Batch Processing Issues: Gemini tends to use scripts for batch processing. If requirements are not precisely defined, it is very easy for it to mess up the project. If the model decides to use a script, it is often quite resistant to subsequent requests to switch to "process one by one."

Agent Context Management Recommendations

• Necessity of Rule Files: Be sure to create rule files like AGENTS.md, otherwise the Agent will repeatedly make the same mistakes.
• WSL Environment Limitations: When using Copilot CLI in WSL, if a PowerShell script is required, the Agent will not automatically convert it to bash, causing the process to break. It is recommended to prioritize using the VS Code extension in WSL environments.
• Long Conversation Lapses: The model's ability to extract information from the beginning of a long conversation will decline. It is recommended to explicitly restate key sections or limit the scope of modifications before regenerating.

---

Change Log

• 2026-03-07 Initial document creation.
• 2026-03-13
  • Added explanations regarding Google's adjustment of Antigravity quota mechanisms on 2026-03-12.
  • Added issues and recommendations regarding Gemini's unstable output integrity.
  • Added limitations of Copilot CLI in WSL environments and rule file reading issues.