Agentic Engineering Principles

1. Define the Goal, Not the Steps

• Don't: "Open file X, find function Y, change line Z"
• Do: "Fix the bug where users can't log in with special characters"

2. Provide Context, Not Instructions

• Don't: "Use React hooks"
• Do: "This project uses React 18 with TypeScript, see CLAUDE.md for conventions"

3. Verify, Don't Trust

• Always have a way to validate agent output
• Tests, type checking, linting, human review
• "Trust but verify" — Compound Engineering principle

4. Iterate in Loops

• Agent tries → You review → Provide feedback → Agent tries again
• This is faster than trying to specify everything upfront

5. Invest in Infrastructure

• Good CLAUDE.md = better agent performance
• Good tests = safer agent autonomy
• Good CI/CD = faster feedback loops

6. Use the "Best of N" Pattern

Generate multiple versions, cherry-pick the best:

You: "Generate 3 different approaches to this authentication flow.
     For each, explain the tradeoffs."

Then:

• Review all versions
• Pick the best one, or combine elements from multiple
• Refine the chosen approach

Why it works: AI generates diverse solutions. You bring judgment to select.

7. Branch First, Experiment Freely

Always create a git branch before major AI-assisted changes:

git checkout -b feature/ai-experiment

• AI can try, fail, and retry without risk
• Easy to discard bad approaches
• Compare AI output against your baseline
• Merge only what works

Prompt Patterns

Effective patterns for getting better results from Claude Code.

Pattern 1: The Specification Prompt

Front-load all context, then state the task:

Context:
- This is a React 18 app with TypeScript
- We use Tailwind for styling
- Components go in /src/components
- We follow the existing Button component pattern

Task: Create a Modal component that:
- Accepts title, children, and onClose props
- Has animated entrance/exit
- Closes on backdrop click and Escape key
- Is accessible (focus trap, aria labels)

Pattern 2: The Critique Prompt

Have Claude review its own work:

You: "Now review the code you just wrote. Check for:
     - Security vulnerabilities
     - Edge cases not handled
     - Performance issues
     - Accessibility problems

     Then fix any issues you find."

Pattern 3: The Three Experts

Get multiple perspectives on design decisions:

You: "Approach this problem as three experts would:
     1. A senior security engineer
     2. A performance optimization specialist
     3. A UX-focused developer

     What would each prioritize? Where do they disagree?"

Pattern 4: The Incremental Build

For complex features, build in stages:

Step 1: "Create the data model and types"
Step 2: "Add the API layer"
Step 3: "Build the UI components"
Step 4: "Connect everything and add error handling"
Step 5: "Write tests for the critical paths"

Pattern 5: The Constraint Prompt

Limit scope to get focused output:

You: "Implement ONLY the login form. Do not:
     - Add routing
     - Create the API client
     - Set up state management

     I'll handle those separately."