Agentic Engineering

Operate as an agentic engineer using eval-first execution, decomposition, and cost-aware model routing. Adapted from everything-claude-code by @affaan-m (MIT).

Quick Start

1. Define completion criteria — write acceptance criteria and success metrics before execution
2. Create baseline evals — write capability and regression tests that capture current state
3. Decompose work — break into 15-minute units, each independently verifiable with a single dominant risk
4. Route models by complexity — Haiku for narrow tasks, Sonnet for implementation, Opus for architecture
5. Run post-implementation evals — measure deltas, confirm no regressions

Key Concepts

• Eval-first execution: Run tests before coding; measure against known baseline; catch regressions early
• 15-minute unit rule: Each task should have one clear risk, one verifiable outcome, be completable in ~15 minutes
• Model tier matching: Complexity determines model — don't overpay for simple tasks, don't underpay for hard ones
• Review focus: Prioritize invariants, error boundaries, security, coupling — not style (automation handles that)
• Session strategy: Continue for coupled units; reset after major phase transitions; compact at milestones