AI Native Notes
Back to Gemini CLI
Gemini CLIIntermediate3 min read

Gemini Task Decomposition — Mastering Complex Implementation

How to leverage Gemini CLI's 1M token context while maintaining precision through strategic task breakdown and execution.

task-decompositionplan-modechecklistverificationworkflow

Official References: Gemini CLI Workflow · Sub-agents · GEMINI.md Guide

Context Isn't Everything

Gemini CLI's 1-million-token context is excellent for architectural understanding, but precise modification of thousands of lines of code is a different challenge. If a task scope is too broad, models are prone to hallucination or missing subtle constraints.

The secret to success is: "Understand globally, execute locally."

Step 1: Strategy with Plan Mode (/plan)

Before starting any implementation, enter read-only mode via the /plan command. The goal here is not to fix code, but to perfectly define "what needs to be fixed."

Strategy Question List

  • Where is the core logic for the current implementation?
  • What dependencies are affected by this change?
  • What test cases will verify success?
  • What are the existing architectural patterns or naming conventions?

Step 2: Create and Agree on a Checklist

Ask Gemini to create a step-by-step execution plan based on its investigation.

User: "Based on your investigation, create an execution plan with no more than 5 steps. Include verification methods (like tests) for each step."

Checklist Example:

  1. Add new authorization logic to src/lib/auth.ts
    • Verify: Run npm test auth.test.ts
  2. Apply middleware to API route src/app/api/data/route.ts
    • Verify: Confirm 403 response for unauthorized access via curl
  3. Update state in frontend component LoginButton.tsx
    • Verify: Visual check in Storybook or local dev server
  4. Add error message keys to all translation files (messages/*.json)
    • Verify: Run check-locale-parity script

Step 3: Progressive Execution (Surgical Edits)

Don't let the agent modify all files at once. Verify after each completed step.

  • Use replace: Encourage the agent to use replace with precise old_string and new_string instead of overwriting entire files.
  • Maintain Session Health: If the session gets too long and quality drops, save progress with /memory add and start a fresh session.

Step 4: Delegate "Local Tasks" to Sub-agents

While the main agent holds the global context, delegate repetitive or local tasks to sub-agents.

  • generalist: "Update license headers in these 10 files."
  • code-review: "Review these modified files for potential security vulnerabilities."

This keeps the main agent's context window clean while progressing in parallel.

Step 5: Final Verification and Memory Persistence

Once finished, run the full build and tests to ensure no side effects. If successful, record repository-specific lessons for your future self.

/memory add "API error handling must always go through src/lib/errors.ts in this project."

Pro Tips for Task Decomposition

  • Start with Failure: When fixing bugs, ask Gemini to write a "failing test case" first. Seeing the test fail before applying the fix is the safest way to develop.
  • No Guessing: If Gemini starts guessing file paths or library versions, stop immediately and ask it to verify facts using grep_search or web_fetch.
  • Step-by-Step Prompting: Instructing "Let's start with the first step" instead of "Do everything at once" can improve accuracy by over 30%.

Next to Read

Connected Guides