security-review

Context Ranker

Use this skill to reduce token waste by reading fewer files, earlier.

Core Principle

Start from the smallest evidence that can confirm/deny a hypothesis.

Workflow

1. Start with the workspace map

   • Read .ai-context/workspace-map.json (first ~200 lines).
   • Use summary.line_guide to jump directly to file metadata.

2. Generate candidates

   • Prefer files explicitly named by the user.
   • Otherwise, shortlist by keyword matching in filenames and directories.

3. Rank (Top-K)

Score files by:

• name match (strong)
• directory relevance (medium)
• description/type match (medium)
• size penalty (prefer smaller)

Default K:

• simple task: 1–3 files
• moderate: 3–5 files
• complex: 5–8 files (only with strong justification)

4. Read surgically
   • Start with the smallest relevant range.
   • Expand only when needed.

Output Format

## Context Ranking

Top files to read next:
1. path/to/file.ext — why
2. path/to/file.ext — why
3. path/to/file.ext — why

Anti-Patterns

• Reading many files before forming a hypothesis.
• Searching broadly before consulting the workspace map.
• Re-reading the same large file multiple times.