smart-review

Smart Review — Injection Pattern Example

Demonstrates the Injection pattern (L13): a Skill that dynamically loads Agent system prompts based on context, rather than always running all agents.

Goal

Unlike /full-review (always dispatches 3 agents in parallel), Smart Review analyzes changed files and dispatches only relevant agents. If only Java files changed, skip the doc-explorer. If only config files changed, run only security-auditor.

How It Works (Injection Pattern)

Smart Review Skill
    │
    ├── Analyze changed files
    │       │
    │       └── .java files? → Inject code-reviewer system prompt
    │       └── .yml/.xml/.env? → Inject security-auditor system prompt
    │       └── .md/.rst? → Inject doc-explorer system prompt
    │       └── Mixed? → Inject multiple (parallel dispatch)
    │
    └── Dispatch selected agents → Collect → Consolidate

Execution

Step 1: Detect Changed Files

git diff --name-only HEAD 2>/dev/null || echo "ALL_FILES"

Step 2: Select Agents Based on File Types

File Pattern	Agent(s) Selected
*.java, *.py, *.ts, *.go	code-reviewer
*.yml, *.yaml, *.xml, .env*, pom.xml, package.json	security-auditor
*.md, *.rst, docs/**	doc-explorer
src/test/**, tests/**	test-runner

Step 3: Inject Agent Prompts

For each selected agent, load its system prompt from ~/.claude/agents/<name>.md and inject it into the review context. This is the Injection pattern — the Skill doesn't duplicate agent definitions, it references them.

Step 4: Parallel Dispatch + Consolidate

Same as /full-review Phase 2-4, but only for selected agents.

Benefits Over /full-review

Aspect	/full-review	Smart Review
Agents dispatched	Always 3	1-3 based on context
Token cost	Higher (3x agent overhead)	Lower (only relevant agents)
Speed	~3 agent runs	~1-2 agent runs
Noise	May flag irrelevant issues	Focused on relevant checks