Overview

Previous Post: #4 — Marketplace Stabilization and v0.3.0 Release

In this #5 installment, the core /harnesskit:architect skill was added across 10 commits and v0.4.0 was released. The architect skill designs multi-agent teams for complex projects, backed by a reference guide covering 6 agent design patterns and orchestrator templates. Additionally, /apply now auto-registers custom agents, hooks, and skills in CLAUDE.md.

Starting from Competitive Analysis

Early in the session, HarnessKit was compared against a competing plugin (revfactory/harness). After mapping out the feature scope, approach, and differentiators of both plugins, gaps in HarnessKit were identified. The biggest gap turned out to be “multi-agent team design for complex projects” — this became the starting point for the architect skill.

/harnesskit:architect — Agent Team Design Skill

Concept

/harnesskit:architect analyzes complex projects and designs multi-agent team structures. It examines the project’s tech stack, scale, and requirements, then recommends appropriate agent compositions and orchestration patterns.

graph TD
    A["User invokes <br/> /harnesskit:architect"] --> B["Project Analysis <br/> Tech stack, scale, requirements"]
    B --> C["Pattern Selection <br/> Match from 6 patterns"]
    C --> D["Team Composition <br/> Assign agents by role"]
    D --> E["Orchestrator Generation <br/> Workflows + error handling"]
    E --> F["Auto-register <br/> in CLAUDE.md"]

Implementation

First, the command registration (/harnesskit:architect) was added to enable autocomplete. Then the skill itself was implemented, with the orchestrator agent enhanced with concrete workflows and error handling logic. A test suite was also written to verify consistency between the skill and reference documents.

Agent Design Patterns Reference

Six design patterns referenced by the architect skill were documented as a reference guide.

graph LR
    subgraph Patterns["6 Design Patterns"]
        P1["Sequential <br/> Pipeline"]
        P2["Parallel <br/> Fan-out"]
        P3["Supervisor <br/> Delegation"]
        P4["Peer Review <br/> Validation"]
        P5["Specialist <br/> Router"]
        P6["Iterative <br/> Refinement"]
    end
    A["architect skill"] --> Patterns
    Patterns --> B["Orchestrator <br/> Templates"]

Each pattern specifies suitable project types, agent configurations, and communication methods. A separate orchestrator templates document was also created, providing concrete implementation templates for each pattern.

CLAUDE.md Auto-Registration — Evolution of /apply

Problem

After creating custom agents, hooks, or skills, manually registering them in CLAUDE.md was tedious and error-prone. If registration was missed, Claude Code would not recognize the agent or hook’s existence.

Solution

Auto-registration was added to /harnesskit:apply. When /apply applies improvement proposals, it detects newly created agents, hooks, and skills, and automatically registers them in the appropriate section of CLAUDE.md.

graph TD
    A["/apply runs"] --> B{"New agents <br/> hooks / skills detected?"}
    B -->|Yes| C["Parse CLAUDE.md"]
    C --> D["Add entries to <br/> relevant section"]
    D --> E["Show registration <br/> results to user"]
    B -->|No| F["Keep existing behavior"]

v0.4.0 Release

The plugin.json version was bumped to 0.4.0, along with adding a homepage URL, author URL, and agent-related keywords. Rich metadata improves discoverability in marketplace search results.

feature_list.json Population

The complete feature inventory of HarnessKit was systematically organized into feature_list.json with a save implementation. This file serves as shared reference data across multiple skills — progress tracking in /harnesskit:status, feature analysis in /harnesskit:insights, and more.

Commit Log

Insights

From competitive analysis to building the architect skill, the core of this session was “finding gaps and filling them.” Multi-agent orchestration is conceptually simple, but practical implementation cascades into design pattern classification, template documentation, and auto-registration. The auto-registration feature in particular fundamentally solves the problem of “creating a tool but forgetting to register it.” Making a tool register itself — that is the essence of good DX (Developer Experience).