News and updates about H·AI·K·U and structured human-AI collaboration.
Anthropic's Dynamic Workflows fan tens-to-hundreds of subagents across one giant task. H·AI·K·U governs the whole feature lifecycle, over and over. Different layers — and they nest.
By Jason Waldrip
H·AI·K·U closes a structural loop where the engine learns from running. The agent writes notes per stage, synthesizes them at intent close, and lands corrections back in the cascade.
By Jason Waldrip
Your CI is green, every unit test passes, every reviewer signed off — and the page is still broken. H·AI·K·U's runtime verifiers open a real browser and look.
By Jason Waldrip
Four posts comparing H·AI·K·U to Anthropic's harness writeup, GSD, GSTACK, and Superpowers. This is the anchor — where we landed on the methodology, who we owe, and what we're actually going for.
By Jason Waldrip
A tighter pass on the Superpowers comparison in the same side-by-side format as the rest of the series. We already wrote the long version; this one's for scanning.
By Jason Waldrip
GSTACK models a 23-person startup team with role-by-role instructions. H·AI·K·U models a workflow engine that emits one action at a time. Same intuition, different architectures.
By Jason Waldrip
GSD-2 and H·AI·K·U both attack context rot. GSD breaks work into atomic plans run in fresh windows. We never let the agent hold the plan in the first place. Same problem, opposite ends.
By Jason Waldrip
Anthropic published a post on harness design for long-running agent apps. We've been building one for a month. Most of it is the same picture. The places we diverge are the interesting part.
By Jason Waldrip
Most agent harnesses are giant skills the model tries to execute end-to-end. H·AI·K·U has always treated workflow as software. v4 is where that lands.
By Jason Waldrip
An agent can have twenty useful slash commands installed and use none of them. Skill alignment plumbs that registry into the elaborate phase so units carry the picks forward.
By Jason Waldrip
H·AI·K·U used to assume the agent owned every file in the intent dir. Drift detection, audit logs, and a human-write MCP tool make the workshop honest.
H·AI·K·U promised that downstream stages execute what upstream stages designed. The promise was a vibe. Three new engine gates make it a contract.
Jesse Vincent's Superpowers and H·AI·K·U look like the same kind of thing and aren't. A field report on where they overlap, where they genuinely differ, and what we're actually aiming for.
By Jason Waldrip
Cascading model selection lets every unit pick the Claude model that fits its complexity — unit, hat, stage, studio, in that order. Cost goes down. Quality goes up. Both.
By Jason Waldrip
Splitting the one-size-fits-all software studio into four product-family lifecycles — and why forcing a library through a UX design sprint was telling us something we didn't want to hear.
Discrete mode now creates a branch per stage. Each stage becomes its own PR. Go-backs merge forward instead of blowing away later work. Hybrid mode picks a threshold and consolidates from there.
By Jason Waldrip
Why H·AI·K·U makes the agent dumb on purpose, keeps state on disk, and only lets the user enter the loop at two seams.
By Jason Waldrip
Quality gates now run before adversarial review agents. Tests, lint, and typecheck fail fast, cheap, and loud — so expensive review cycles never get spent on code that doesn't compile.
By Jason Waldrip
The review UI used to live on localhost. Now it lives on haikumethod.ai, reaches back to your machine over a tunnel, and lets you approve a stage from your phone at lunch.
By Jason Waldrip
H·AI·K·U was born at The Bushido Collective and is now owned and maintained by GigSmart. Here's how we use it — and why we took ownership.
By Jason Waldrip
From AI-DLC to H·AI·K·U — a methodology evolution from software-specific lifecycle to domain-agnostic work orchestration.
When AI handles implementation, the value of a human shifts from domain expertise to systems thinking. The agnostic builder doesn't specialize — they see the whole board.
H·AI·K·U now includes a design direction system and persistent project knowledge layer, making visual design and domain understanding part of the methodology — not an afterthought.
H·AI·K·U gains first-class integration with six design tools and a persistent knowledge system that gives agents institutional memory across features.
H·AI·K·U now supports typed iteration through disciplinary lenses — design, product, dev — where each pass shapes how hats behave and which workflows are available.
Two new operating modes for H·AI·K·U — quick mode for trivial tasks with full hat discipline, and autopilot for autonomous feature delivery with strategic human checkpoints.
H·AI·K·U now supports reverse-engineering existing codebases into structured intent artifacts, enabling brownfield systems to participate in the full lifecycle.
Anthropic's engineering blog defines 'harness' as orchestration scaffolding for long-running AI agent work. H·AI·K·U fits that definition — and extends it.
The dark factory isn't a system you build. It's a knob you turn. H·AI·K·U treats full autonomy as a point-in-time decision, not an architectural commitment.
H·AI·K·U's construction loop now leverages Claude Code's Agent Teams, turning each unit of work into an independent teammate with its own context, worktree, and permission model.
A methodology for iterative AI-driven development with hat-based workflows