H·AI·K·U Guide

H·AI·K·U Guide

H·AI·K·U = Human + AI Knowledge Unification. A lifecycle framework for structured AI-assisted work.

How It Works

All commands are MCP prompts on the haiku server. Invoke them as /haiku:new, /haiku:resume, etc. Do NOT use the Skill tool for haiku commands — they are MCP prompts, not skills.

The orchestrator tool haiku_run_next { intent } drives everything. It returns an action + stage_metadata. Follow the action, then call it again. Repeat.

Hierarchy

Studio → Stage → Unit → Bolt

• Studio — lifecycle template (e.g., software, ideation). Defines stages.
• Stage — phase within a studio (e.g., inception, design, development). Has hats, review gates, unit types.
• Unit — discrete piece of work within a stage. Has completion criteria and a dependency DAG.
• Bolt — iteration cycle within a unit. Increment when criteria aren't met.
• Hat — behavioral role scoped to a stage (e.g., architect, builder, reviewer).

Stage Scope (Critical)

Every haiku_run_next response includes stage_metadata — the stage's description, allowed unit_types, and STAGE.md body. ALL work MUST stay within the stage's scope. Do not produce outputs belonging to other stages.

Phase Distinction (Critical)

Elaborate and Execute are fundamentally different:

• Elaborate = broad research + unit definition. No hat rotation. Produce discovery artifacts (knowledge/) and unit specs (scope, criteria, dependencies). Do NOT produce the actual deliverables — units are instructions for focused work, not the work itself. One adversarial review before the human gate.

• Execute = focused work through hats. Each unit runs through the stage's hat sequence (e.g., architect → builder → reviewer). Each hat is a separate subagent. Hats ensure completeness via different perspectives. This is where deliverables are produced and written to stages/{stage}/artifacts/.

Elaboration produces the PLAN. Execution produces the WORK.

Key MCP Tools

Tool	Purpose
haiku_run_next	FSM driver — reads state, performs mutation, returns next action
haiku_go_back	Go back to a previous stage or phase (human-initiated)
haiku_intent_list	List all intents
haiku_intent_get	Read an intent field
haiku_stage_get	Read a stage field
haiku_unit_start	Start a unit
haiku_unit_advance_hat	Move to next hat
haiku_unit_complete	Complete a unit
haiku_unit_increment_bolt	New iteration cycle
ask_user_visual_question	Open question form (blocks until user responds)
pick_design_direction	Open design direction picker
haiku_studio_list	List available studios

Review gates are handled automatically by haiku_run_next — it opens the review UI, blocks until the user decides, and returns the outcome. Do NOT call open_review directly.

Commands (MCP Prompts)

Core: new, run, refine, review, reflect Supporting: autopilot, composite, setup, migrate, scaffold, operate, triggers, dashboard, backlog Niche: adopt, quick, seed, ideate, pressure-testing, capacity, release-notes