agent-builder

Guide Claude Code users in authoring reliable sub-agents with structured frontmatter and verifiers.

15|2|Updated May 23, 2026
One-click install
npx skills add https://github.com/VKirill/antigravity-for-claude-code --skill agent-builder-vkirill
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: agent-builder
Source: https://github.com/VKirill/antigravity-for-claude-code/tree/main/skills/agent-builder
Command: npx skills add https://github.com/VKirill/antigravity-for-claude-code --skill agent-builder-vkirill

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires bash, and includes references (resource) and scripts (resource) components.

What problem does it solve?

Agent-builder helps you decide when to use Claude Code skills versus sub-agents, and gives you a proven structure for authoring sub-agents that delegate reliably (planner, verifier, and tool-restricted operators) without creating coordination failures.

Core Features & Use Cases

  • Sub-agent vs skill decision guidance: choose the right abstraction for context isolation, parallelization, and hard tool restrictions.
  • Proven planner pattern: produce SPEC.md + checklist + file budgets, then stop (main implements).
  • Proven verifier pattern: run complete checks with explicit MUST instructions to avoid Early Victory.
  • Tool-restricted operator pattern: enforce guarantees with PreToolUse hooks (e.g., read-only DB querying).
  • Ready-to-use agents and templates for Kirill’s Claude Code workflow (dev-orchestrator, feature-planner, test-verifier, security-verifier, payments-verifier, db-reader).

Quick Start

Use agent-builder to design a new sub-agent by reading the relevant references, then adapt one of the provided templates into a new file under ~/.claude/agents/ with a concrete description and the appropriate tools and permissionMode.

Frequently Asked Questions about agent-builder

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I design Claude Code sub-agents that delegate reliably without coordination failures?

Claude Code sub-agents prevent coordination failures by using a planner pattern that produces a specification and checklist, then stops. They require structured frontmatter and complete-check verifiers to avoid Early Victory during implementation.

When should I use Claude Code skills versus sub-agents for a task?

Choose Claude Code sub-agents over skills when you need context isolation, parallelization, or hard tool restrictions. Skills are better suited for tasks not requiring strict operator separation or blackbox verifier checks.

What is the best way to enforce tool restrictions on Claude Code sub-agents?

The best way to enforce tool restrictions on Claude Code sub-agents is by using PreToolUse hooks. This pattern guarantees safe operations, such as restricting an operator to read-only database querying within the workflow.

How do I prevent Early Victory when running sub-agent code verification?

To prevent Early Victory during sub-agent code verification, implement a proven verifier pattern that runs complete checks with explicit MUST instructions. This ensures the blackbox verifier confirms full task completion.

Does the planner pattern require a specific specification format for Claude Code agents?

The planner pattern requires a SPEC.md file accompanied by a checklist and file budgets. The sub-agent generates this specification and then stops, leaving the main implementation to execute the planned steps.

Can I use bash scripts to set up PreToolUse hooks for read-only database querying?

Yes, you can use bash scripts to configure PreToolUse hooks for read-only database querying. This setup provides a tool-restricted operator pattern that enforces safe operational guarantees within Claude Code.