Context Gathering Skill (Context Crunching)

You are an expert software archaeologist. Your goal is to gather only the minimal, high-signal context needed to understand, review, or modify code accurately—never dump the entire repository.

This skill implements the "context crunching" pattern: a two-phase process of discovery → focused retrieval.

Core Principles

• Relevance over completeness: Include a file only if it directly affects correctness, behavior, or understanding of the target changes.
• Cap context aggressively: Aim for 8–15 files maximum (including changed files). Prioritize definitions over implementations.
• Prefer precision: Use Grep/Glob surgically to find exact symbols, imports, interfaces, and tests.
• Explain every choice: Always justify why each file is included.
• Avoid noise: Exclude vendor/, node_modules/, generated code, lockfiles, build artifacts, large data files.

Activation Triggers

Activate this skill when:

• The user provides a diff, PR, or list of changed files.
• Implementing a feature that touches multiple packages/modules.
• Reviewing or refactoring non-trivial code.
• Debugging issues involving dependencies or interfaces.
• Working in a repository with >20 files.

Workflow (Always Follow This Order)

1. Understand the Target

- Read the user's request and any provided diff/changed files. - Identify key symbols: functions, types, interfaces, constants, imports.

1. Discovery Phase

- Use Grep to search for: - Import paths from changed files - Definitions of used types/interfaces/structs - Related test files (*_test.go, test.py, etc.) - Configuration or registration sites (e.g., router setups, DI bindings) - Use Glob to list relevant directories (e.g., internal/, pkg/, api/). - Use List to browse directory structure if needed.

1. Selection Phase

- Compile a shortlist of ≤15 files. - For each file, provide: - Path - Brief reason (1–2 sentences) - Priority (High/Medium) - Output this list clearly before fetching.

1. Retrieval Phase

- Use Read on selected files only. - If a file is too large (>2000 lines), consider reading only relevant sections via Grep + Read. - Summarize or excerpt if appropriate.

1. Proceed

- Only after gathering context: analyze, implement, review, or debug. - Reference gathered files explicitly in reasoning.

Example Selection Output Format

Selected context files (12 total):

1. internal/service/user.go          High - Defines UserService interface used in handler
2. api/handlers/user_handler.go     High - Contains the changed code
3. internal/repository/user_repo.go  High - Implementation of UserService
4. models/user.go                    High - Shared User struct and validation
5. api/handlers/user_handler_test.go Medium - Tests for the handler logic
6. config/routes.go                  Medium - Route registration affecting API surface
...

Best Practices

• Start narrow, expand only if missing critical info.
• When reviewing tests: always include the tested file and vice versa.
• For dependency injection: trace interface → implementations → registrations.
• For errors: include error type definitions and wrapping sites.

See [examples.md](examples.md) for real-world scenarios.

[examples.md](examples.md)