| name | ai-dlc-mode-selection |
| description | Use when deciding between HITL, OHOTL, and AHOTL modes in AI-DLC workflows. Covers decision frameworks for human involvement levels and mode transitions. |
| allowed-tools | ["Read","Grep","Glob"] |
AI-DLC Mode Selection
AI-DLC supports three modes of human-AI collaboration. Choosing the right mode for each phase of work is critical for balancing productivity, quality, and control.
The Three Modes
HITL - Human In The Loop
Human actively participates in every decision.
Characteristics:
- Every action reviewed by human
- Synchronous collaboration
- Human makes final decisions
- AI proposes, human disposes
When to Use:
- Defining requirements (elaboration)
- Making architectural decisions
- Reviewing security-sensitive changes
- Course corrections when AI is off track
Examples:
User: "Add user authentication"
AI: "What authentication method? OAuth, email/password, or both?"
User: "Start with email/password, we'll add OAuth later"
AI: "Should we support 'remember me'?"
User: "Yes, 30-day sessions"
OHOTL - Occasional Human Over The Loop
Human sets direction, AI operates with periodic checkpoints.
Characteristics:
- AI works autonomously on defined tasks
- Human reviews at milestones
- Human intervenes when stuck or for approval
- Balance of autonomy and oversight
When to Use:
- Building well-defined features
- Tasks with clear completion criteria
- When backpressure provides quality gates
- Medium-complexity work
Examples:
User: "Implement the login form based on these criteria"
AI: [Works autonomously]
AI: "Login form complete. Tests passing. Ready for review."
User: "Looks good, continue to the API integration"
AHOTL - Autonomous Human Over The Loop
AI operates with minimal human involvement.
Characteristics:
- AI makes most decisions independently
- Human reviews only at completion or on exception
- Requires very clear criteria and robust backpressure
- Maximum autonomy
When to Use:
- Well-defined, routine tasks
- Tasks with comprehensive test coverage
- When all edge cases are known
- Low-risk changes
Examples:
User: "Implement all the CRUD endpoints for the User model"
AI: [Completes multiple iterations autonomously]
AI: "All endpoints implemented. 47 tests passing. PR ready."
User: "Merged."
Mode Selection Framework
Decision Matrix
| Factor | HITL | OHOTL | AHOTL |
|---|
| Requirements clarity | Low | Medium | High |
| Risk level | High | Medium | Low |
| Test coverage | Low | Medium | High |
| Domain familiarity | Low | Medium | High |
| Reversibility | Difficult | Moderate | Easy |
Questions to Ask
-
How clear are the requirements?
- Vague → HITL
- Mostly clear → OHOTL
- Crystal clear → AHOTL
-
What's the risk of mistakes?
- Security/data loss → HITL
- User-facing bugs → OHOTL
- Internal tooling → AHOTL
-
How good is test coverage?
- No tests → HITL
- Some tests → OHOTL
- Comprehensive tests → AHOTL
-
How familiar is the domain?
- New/complex domain → HITL
- Familiar patterns → OHOTL
- Routine work → AHOTL
-
How reversible are changes?
- Database migrations → HITL
- API changes → OHOTL
- Internal refactoring → AHOTL
Mode by Phase
Default Workflow Modes
| Phase | Default Mode | Rationale |
|---|
| Elaboration | HITL | Requires human input for requirements |
| Planning | HITL | Human should validate approach |
| Building | OHOTL | Autonomous with backpressure |
| Review | HITL | Human verification before completion |
Mode Overrides
You can override defaults in .ai-dlc/hats.yml:
hats:
builder:
mode: AHOTL
instructions: |
Work autonomously. Only stop if blocked.
Transitioning Between Modes
Upgrading Autonomy (HITL → OHOTL → AHOTL)
When to upgrade:
- Requirements have stabilized
- Test coverage is comprehensive
- Pattern is established
- Human has built trust
Example:
Session 1 (HITL): Define auth requirements together
Session 2 (HITL): Review initial implementation
Session 3 (OHOTL): AI implements remaining endpoints
Session 4 (AHOTL): AI handles routine CRUD operations
Downgrading Autonomy (AHOTL → OHOTL → HITL)
When to downgrade:
- Unexpected complexity discovered
- AI making repeated mistakes
- Security concerns arise
- Requirements changed
Example:
AI operating in AHOTL...
AI: "I'm stuck on edge case X. Need clarification."
→ Downgrade to HITL for this issue
→ Resume OHOTL once resolved
Mode Indicators
Signs You're in the Wrong Mode
Too much autonomy (should downgrade):
- Repeated mistakes on similar issues
- Misunderstanding requirements
- Missing edge cases
- User frequently correcting course
Too little autonomy (should upgrade):
- User rubber-stamping every decision
- Routine, repetitive work
- Comprehensive test coverage exists
- AI consistently making good decisions
Calibration Questions
Ask periodically:
- "Am I making decisions the human should make?"
- "Am I asking for approval on routine choices?"
- "Are my autonomous decisions causing rework?"
- "Is the human adding value at this checkpoint?"
Mode-Specific Behaviors
In HITL Mode
- Ask before every significant decision
- Present options with trade-offs
- Wait for explicit approval
- Document decisions with rationale
In OHOTL Mode
- Make routine decisions autonomously
- Check in at milestones
- Ask when genuinely uncertain
- Save progress frequently (han keep)
In AHOTL Mode
- Make all decisions within criteria bounds
- Only interrupt for true blockers
- Log decisions for later review
- Complete full task before seeking feedback
Backpressure by Mode
HITL Backpressure
Human IS the backpressure:
- Every change reviewed
- Human catches issues immediately
- No automated gates needed
OHOTL Backpressure
Mix of automated and human:
- Automated: tests, lint, types
- Human: milestone reviews, PR approval
- AI operates freely within automated bounds
AHOTL Backpressure
Fully automated:
- Comprehensive test suite
- Strict type checking
- Automated code review tools
- CI/CD pipeline as final gate
Examples
Example 1: Security Feature
phase: authentication
mode: HITL
reason: |
Authentication has security implications.
Every decision needs human review.
Example 2: UI Component
phase: component_library
mode: OHOTL
reason: |
Design system is established.
AI implements, human reviews at milestones.
Example 3: Data Migration Script
phase: migration
mode: HITL
reason: |
Database changes are difficult to reverse.
Human must verify each step.
Example 4: Unit Tests
phase: test_writing
mode: AHOTL
reason: |
Tests are additive and easily reversible.
Existing tests validate correctness.
Summary
| Mode | Human Involvement | Use When |
|---|
| HITL | Every decision | High risk, unclear requirements |
| OHOTL | At milestones | Medium risk, clear criteria |
| AHOTL | At completion | Low risk, comprehensive tests |
Default rule: Start with HITL for new work, upgrade autonomy as trust builds and tests accumulate.
