stegoro

stegoro

Automatically generate steering policies that drive purpose-built AI agents through structured, quality-assured workflows.

English • 日本語

Why

Building a purpose-built AI agent is hard. Without a structured policy, agents produce inconsistent outputs, skip critical validation steps, and lack quality assurance. Writing these policies by hand is time-consuming and error-prone.

stegoro solves this by automating the entire policy creation process — from understanding the agent's purpose to generating a complete, quality-calibrated policy set.

What

A meta-agent for Claude Code that generates steering policies for purpose-built AI agents. It produces a complete set of rule files that govern an agent's behavior through structured, multi-phase workflows.

Features

• 4-Phase Workflow — Discovery, Design, Generation, and Refinement phases ensure thorough policy creation
• Adaptive Depth — Automatically adjusts research and design depth based on domain complexity (Minimal / Standard / Comprehensive)
• 4 Agent Types — Supports Process, Task, Analytical, and Hybrid agent patterns
• AI-DLC Quality Assurance — Validates output against 11 quality dimensions including adaptive workflow, mandatory checkpoints, audit trail, and overconfidence prevention
• Interactive Dialogue — Structured question-and-answer flow with approval gates at every critical decision point
• Session Continuity — State tracking enables pause and resume across sessions
• Complete Audit Trail — ISO 8601 timestamped logging of all decisions and user inputs

Steering Policies as a User Harness

The generated policies function as a user harness — a structured control layer that wraps around the AI agent to ensure it executes complex workflows predictably and under human supervision.

Without a harness, an AI agent operates freely: it may skip validation steps, make autonomous decisions without approval, or deviate from the intended workflow. Steering policies solve this by defining explicit rules the agent must follow at every step.

How the harness controls workflow:

Control Mechanism	How It Works
Approval Gates	The agent cannot advance to the next phase or stage without explicit user confirmation
Phased Execution	Work is divided into discrete stages with defined inputs, outputs, and completion criteria
Mandatory Checkpoints	Critical decisions (scope, architecture, quality) always require user review before proceeding
Overconfidence Prevention	The agent is instructed to ask rather than assume when uncertain, preventing silent errors
Audit Trail	Every decision and user input is logged with ISO 8601 timestamps for accountability

The core-workflow.md acts as the master orchestrator, routing the agent through each stage in order. The rule-details/ files provide deep, domain-specific instructions that the agent reads on demand — keeping context focused and instructions precise. Together they ensure the agent behaves as a predictable, controllable system rather than a free-form AI.

Packaging as an Agent Skill

The generated user harness can be packaged as a Claude Code agent skill, turning it into a distributable, installable plugin.

By adding a .claude-plugin/ metadata directory alongside the generated policy files, the harness becomes a plugin that anyone can install and invoke via slash commands — no manual file setup required.

# Install a published agent skill
/plugin marketplace add <org>/<agent>
/plugin install <agent>@<agent>

# Invoke the agent
/<agent>:run

This is the deployment model used by the example agents in the samples/ directory — each includes a .claude-plugin/ directory that makes it publishable to the Claude Code marketplace.

Generated Output