𝗧𝗵𝗲 𝟭𝟬 𝗟𝗶𝗻𝗲 𝗦𝗲𝗰𝗿𝗲𝘁 𝘁𝗼 𝗕𝗲𝘁𝘁𝗲𝗿 𝗔𝗜 𝗣𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲

𝗧𝗵𝗲 𝟭𝟬-𝗟𝗶𝗻𝗲 𝗦𝗲𝗰𝗿𝗲𝘁 𝘁𝗼 𝗕𝗲𝘁𝘁𝗲𝗿 𝗔𝗜 𝗣𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲

Stop writing long instruction files for your AI. Most users waste tokens by adding unnecessary details to their project files.

I tested how the length of a GEMINI.md file affects Gemini CLI performance. Here are the results.

I compared three versions:

• GEMINI_0: Empty file.
• GEMINI_10: 10 lines of core settings (language, indentation, server location).
• GEMINI_100: 100 lines of full background, naming rules, and tool settings.

The results show that 10 lines are just as effective as 100 lines.

The Data: • GEMINI_0: 33.3% accuracy | 66.8s latency • GEMINI_10: 100% accuracy | 16.1s latency • GEMINI_100: 100% accuracy | 20.2s latency

Why 10 lines win:

• Accuracy stays at 100%.
• It uses 87% fewer tokens than the long version.
• It responds faster.

Rules for writing instructions:

• Do NOT include: Project stories, official documentation, or general coding standards (like PEP8). The AI already knows these.
• DO include: Specific language requirements, project-specific naming rules, custom SSH mappings, or non-standard local ports.
• Use this rule: If you can find the rule on Google, do not put it in your instruction file.

Other ways to save tokens:

• Context Caching: Use this for repetitive long prompts.
• Prompt Compression: Use tools like LLMLingua for long documents.
• RAG: Use vector search instead of stuffing entire files into context.
• Output Control: Use JSON format to reduce response length by 85%.

Summary of strategies:

• Slim system prompts: Best for all CLI and API users.
• Context Caching: Best for repetitive long prompts.
• JSON Output: Best for structured data tasks.

Keep your instructions lean. Only write what the model does not already know.

Source: https://dev.to/jh5_pulse/rang-sheng-huo-chan-sheng-bian-hua-de-geminixiao-shi-yan-men-45ke

Optional learning community: https://t.me/GyaanSetuAi