Human-AI Collaboration Model
In one line: The human owns what and why; the AI owns how — under human validation and review.
What: AI-assisted development is a collaboration between an architect (human) and a senior engineer (AI), not automation. The human decides what to build and why, and validates proposals against requirements, constraints, and architectural coherence. The AI proposes how, implements, and the human reviews.
Why: Two opposite failure modes motivate the split.
AI without judgment. Unsupervised, the AI produces code that is locally correct but globally incoherent — each component works alone, the system doesn't compose. It lacks business context and long-term vision, optimizes for the immediate request, and ignores cross-cutting concerns.
Human without velocity. A solo architect implementing a complex system by hand spends months after the design is done. The architecture is sound but the time-to-market is not. Judgment is irreplaceable; typing speed is a bottleneck.
The model aims to combine accountable judgment with faster drafting, implementation and testing. Verify requirements coverage, missed defects and rework: AI output can be incomplete or inconsistent even when it follows familiar patterns.
Specific Division of Labor:
| Responsibility | Owner | Rationale |
|---|---|---|
| Problem definition | Human | Requires business context the AI does not have |
| Approach selection | Human (with AI proposals) | AI proposes 3+ approaches; human selects based on constraints AI cannot fully evaluate |
| API contract design | Human | Contracts affect other teams and external consumers; judgment-intensive |
| Implementation | AI (under human review) | Pattern-following work where AI excels |
| Test implementation | AI-assisted engineering, with accountable review | Generate cases, then check coverage and oracle independence against approved expectations |
| Code review | Both (two-stage) | AI reviews for spec compliance and patterns; human reviews for architectural coherence |
| Regulatory compliance | Human (with AI support) | Legal interpretation requires human accountability; AI assists with completeness checks |
| Documentation | AI (under human review) | Draft and update claims; reviewers verify their meaning, evidence and audience |
| Refactoring | AI (under human direction) | Propose scoped transformations; regression checks and review assess preserved behavior |
| Architecture decisions | Human | ADRs require judgment about long-term tradeoffs |
Evidence: Commit attribution can record stated contributors, but does not establish who made or approved a decision. Use scoped task, review and approval records to evaluate the division of work and its outcomes.
How: Instructions, retrieval and review support this collaboration; action-boundary controls supply only the enforcement actually installed and tested:
-
Project instructions provide compact pointers to the adopted architecture, commands and constraints. Verify host loading and keep them beneath binding organizational authority.
-
Memory files help retrieve relevant prior context. Check their audience, currency and authority; a link does not prove the host loaded its target.
-
Superpowers skills guide lifecycle sequencing. They can be omitted or misapplied; their presence does not prove design approval or executed verification.
-
Agent-based review (Section 5) provides automated second opinions. The human architect is not the sole quality gate — specialized reviewer agents check for security issues, API consistency, and compliance gaps.
-
Subagent-driven development, when selected and authorized, narrows task context and separates ownership. Verify the context supplied and returned evidence; fresh sessions can still share mistaken assumptions.
For the full agent architecture, see Section 5. For memory system details, see appendix-d-memory.md.