The HAIKU Methodology
A concise overview of the HAIKU framework. For the complete treatment, see the full paper.
The Problem
Most teams rely on ad-hoc prompting — improvised interactions with no persistent structure, no quality enforcement, and no learning loop. This creates predictable failure modes:
Core Principles
Reimagine Rather Than Retrofit
Traditional methodologies were designed for human-driven processes with long iteration cycles. With AI, iteration costs approach zero. HAIKU is built from first principles for this reality — continuous flow with strategic checkpoints rather than discrete phases separated by gates.
Quality Enforcement Over Prescription
Instead of specifying step-by-step procedures, define the constraints that must be satisfied. Let AI determine how to satisfy them. Quality gates reject non-conforming work without dictating approach.
"Don't prescribe how; create gates that reject bad work."
Context Preservation Through Artifacts
AI context windows reset between sessions. HAIKU addresses this through artifact-based persistence: the outputs of each phase serve as structured context for subsequent work. This creates a self-documenting workflow.
Iterative Refinement Through Bolts
Work progresses through bolts — iteration cycles within units. Each bolt produces a reviewable increment. Small cycles with frequent feedback prevent drift and compound learning.
Human Oversight at Strategic Moments
Human judgment remains essential but should be applied where it matters most. Three collaboration modes define the spectrum:
| Mode | Human Role | AI Role |
|---|---|---|
| Supervised | Directs and approves | Proposes, explains, executes on approval |
| Observed | Monitors, intervenes when needed | Executes continuously, accepts redirection |
| Autonomous | Defines boundaries, reviews outcomes | Executes independently within constraints |
Learning Loops
Reflection is not optional. Every completed initiative feeds learnings back. Future elaboration draws on past reflection. Teams that use HAIKU get better at using HAIKU.
The 4-Phase Lifecycle
Terminology
| Term | Definition |
|---|---|
| Intent | The thing being accomplished — the top-level goal or initiative |
| Unit | A discrete piece of work within an intent |
| Bolt | An iteration cycle within a unit that produces a reviewable increment |
| Hat | A behavioral role assumed during execution (e.g., planner, executor, reviewer) |
| Workflow | An ordered sequence of hats defining how a unit progresses |
| Quality Gate | A configurable verification checkpoint that provides backpressure |
| Profile | A domain-specific implementation of HAIKU (e.g., AI-DLC, SWARM) |
| Collaboration Mode | The human-AI interaction pattern: Supervised, Observed, or Autonomous |
For the complete methodology with all details, examples, and formal definitions:
Read the Full Paper