Gemini Context Master

Gemini Context Master

This skill implements senior-level context engineering to maximize token efficiency and maintain long-term project memory.

1. Strategic Compaction

Do not wait for auto-compaction. Manually trigger or suggest compaction at logical task boundaries to prune old exploration data while keeping the "Final Plan" and "Current Progress".

When to Compact:

• After a long research phase and before starting execution.
• After successfully completing a major milestone.
• When the context feels "sluggish" or token count is high (>50k).

Compaction Workflow:

1. Summarize the Current State (what has been done).
2. Carry over the Pending Tasks (what is next).
3. Carry over any Crucial Findings (API keys locations, architectural decisions).
4. Run /compact.

2. Surgical Context Pruning

Avoid reading entire large files. Use targeted tools to keep the context clean.

• Prefer grep_search: Find specific lines instead of reading 1000+ lines.
• Use read_file with line ranges: Only read the relevant function or class.
• Prune Output: If a command produces massive output, only carry over the error or the relevant success message.

3. Persistent Project Memory (GEMINI.md)

Every project should have a GEMINI.md file in the root. This acts as a "Source of Truth" that saves tokens by preventing repetitive questions.

GEMINI.md Structure:

• Tech Stack: Languages, frameworks, versions.
• Conventions: Naming styles, directory structure.
• Manual Overrides: Specific instructions for this project only.

4. Token-Saving Guidelines

• No Chitchat: Keep responses technical and direct.
• Aggregated Tool Calls: Run independent tasks in parallel within one turn.
• Contextual Precedence: Always check GEMINI.md before asking for project details.