AI-First Engineering

Engineering operating model for teams where AI agents generate a large share of implementation output. Adapted from everything-claude-code by @affaan-m (MIT).

Quick Start

1. Invest in planning quality — ambiguous specs cause AI-generated code to fail; write clear acceptance criteria first
2. Raise eval coverage — AI code requires higher test standards; regression coverage mandatory for touched domains
3. Shift review focus — review for behavior, security, data integrity, failure handling; let automation handle style
4. Design agent-friendly architecture — explicit boundaries, stable contracts, typed interfaces, deterministic tests
5. Evaluate hiring signals — decomposition skill, measurable criteria definition, prompt quality, risk control discipline

Key Concepts

• Planning > Speed: Clear specs + good evals trump fast typing. AI can implement fast; humans must specify clearly.
• Automation is the baseline: Style, formatting, lint issues are solved by automation, not review.
• Architecture matters more: Implicit conventions break AI systems; use explicit boundaries and typed interfaces.
• Test coverage is non-negotiable: Generated code needs regression coverage for every touched domain.
• Shared responsibility: AI generates; human reviews for risk (security, data integrity, rollout safety); human refines when needed.

Common Usage

Code review in AI-first teams — focus on:

Behavior regressions: Did the change break existing functionality?
Security assumptions: Input validation, permission checks, sensitive data handling
Data integrity: Constraints, rollback safety, concurrent access
Failure handling: Network calls, database errors, timeouts, degraded modes
Rollout safety: Feature flags, backward compatibility, canary deploy strategy

Architecture for AI teams:

• Explicit boundaries between modules (not implicit conventions)
• Stable contracts (typed interfaces, documented behavior)
• Deterministic tests (no flaky tests — AI can't debug intermittent failures)
• Clear error paths (AI struggles with ambiguous error handling)

Testing standard raise:

• Regression coverage for every touched domain (required, not optional)
• Explicit edge-case assertions (AI may miss corner cases)
• Integration checks for interface boundaries (behavior across module lines)

Hiring Signals for AI-First Engineers

Strong signals:

• Decomposes ambiguous work cleanly → clear, testable units
• Defines measurable acceptance criteria → no scope creep, clear done condition
• Produces high-signal prompts and evals → AI generates better code from better specs
• Enforces risk controls under delivery pressure → doesn't skip security or testing for speed

Weak signals:

• "Move fast and break things" mindset
• Writing code without clear specs or acceptance criteria
• Skipping regression tests to save time
• Vague PR descriptions ("fixed bugs," "refactored stuff")

References

• references/process-shifts.md — detailed planning, evals, review guidance
• references/architecture-guide.md — designing systems for AI code generation
• references/testing-standards.md — regression coverage, edge-case testing, integration checks