Human In Loop

HumanInLoop

Stop vibe coding. Ship software that lasts.

Website • Quick Start • Roadmap • Changelog

What is HumanInLoop?

HumanInLoop is a Claude Code plugin that enforces specification-driven development—ensuring architectural decisions are made by humans before AI writes code.

Instead of letting AI improvise your architecture, you guide it through a structured workflow:

Every step produces artifacts you can review, refine, and approve before moving forward.

The Workflow

Each command produces artifacts you review before the next step. You stay in control.

Command Details

Specify - Create feature specification (DAG-based)

Output: specs/{feature}/spec.md + StrategyGraph JSON

Agents:
- State Analyst — Strategic analysis + graph mechanics; produces briefings, assembles nodes, freezes passes, and constructs domain agent prompts
- Requirements Analyst — Transforms feature requests into structured specs; no implementation details
- Devil's Advocate — Reviews for gaps and ambiguity; asks clarifying questions

Plan - Unified analysis and design planning

Output: specs/{feature}/plan.md + 6 supporting artifacts

Agents:
- Technical Analyst — Produces analysis artifacts (requirements, constraints-and-decisions, NFRs) and design artifacts (data model, API contracts, integration guide)
- Principal Architect — Reviews cross-artifact feasibility after analysis phase (one-time gate)
- Devil's Advocate — Validates completeness, traceability, and cross-artifact consistency

Tasks - Generate implementation tasks

Output: specs/{feature}/tasks.md with TDD cycles

Agents:
- Task Architect — Maps requirements to vertical slices with TDD structure; no implementation
- Devil's Advocate — Ensures coverage, proper ordering, and testable increments

Setup - Create project…