AI Native Notes
Back to Gemini CLI
Gemini CLIAdvanced3 min read

Gemini Effective Prompting — Mastering the 1M Token Context

Move beyond short chat prompts to specialized prompt engineering for the Gemini CLI and its massive 1M token context window.

promptingengineeringcontext-managementxml-tagsefficiency

Core Principle: Having a 1-million-token context window does not mean the model's attention is infinite. The density and structure of your prompt determine the quality of the result.

Chat Prompts vs. Agent Prompts

When talking to ChatGPT or Claude in a browser, a short, abstract prompt like "Make me this" might work. But for an agent modifying your file system directly from the terminal, a completely different approach is needed.

  • Bad Example: "Add a Google login button to the login page."
  • Good Example: "Analyze the structure of /src/components/auth, then add a Google login button applying the same Tailwind style guide as existing buttons. Follow the existing api/auth/ pattern for the API endpoint."

1. Structured Instructions with XML Tags

The Gemini CLI model parses chunks of information much better when they are wrapped in XML tags. Actively use tags when giving long instructions.

<context>
We are currently migrating from the Pages Router to the App Router.
</context>
 
<goal>
Convert `/pages/dashboard.tsx` to `/app/dashboard/page.tsx`.
</goal>
 
<constraints>
- Minimize the use of client components (`"use client"`).
- Perform data fetching in server components.
- Delete the existing `getServerSideProps`.
</constraints>
 
Based on these guidelines, please set up a migration plan in `/plan` mode first.

2. Enforcing Chain of Thought

Directing the agent to go through an analysis and planning phase before modifying code drastically reduces hallucinations.

Prompt Example:

"Before performing the next task, you must: 1) Read 3 related existing files first, 2) Present a list of parts to be modified to me, and 3) Get my approval before executing."

3. Negative Prompting

Clarifying what NOT to do is just as important as what to do. Especially in large refactors, AI often deviates from the original goal to do unnecessary "cleanup."

"Focus solely on the requested bug fix. If you find lint errors or outdated comments in related files, do NOT modify them unless they are within the scope of this specific task."

4. Preventing Context Dilution

Just because you can fill 1 million tokens doesn't mean you should. Mixing irrelevant files can cause the model to borrow wrong variable names or patterns.

  • Explicit Inclusion: gemini "@src/components/button.tsx @src/styles/colors.css Apply color variables"
  • Use Exclusion Patterns: When having it use the grep_search tool, specify: "Always exclude node_modules or dist folders when searching." (It's best to put this as a global rule in GEMINI.md.)

5. Persona Adoption

Assigning a role sharpens the model's response tone and the nature of the code it generates.

"You are a senior security engineer with 10 years of experience. From now on, relentlessly find potential SQL Injection, XSS, and SSRF vulnerabilities in the code I present and write defensive code."

Next to Read

Connected Guides