Agent Skill Builder Kit

For people building their own agent setup: write skills, AGENTS.md and CLAUDE.md files and sub-agents, manage context, and audit instructions for drift.

A setup for AI coding agents, made by Kitloom. One source, generated for every major agent: Claude Code, OpenAI Codex, Cursor, Google Antigravity, Gemini CLI, GitHub Copilot, Windsurf (Devin), opencode and Aider.

Install

1. Download this repository as a ZIP, or use the Check and download button on its Kitloom page, which checks it first.
2. Read the files. Nothing here runs on its own: the kit has no hooks and no install scripts.
3. Copy the folders and files you want into the root of your project. AGENTS.md is the shared instructions file; edit it to fit your project.

What is inside

Skills

• writing-for-agents — Creating or editing skills, AGENTS.md or CLAUDE.md so the text is written for an agent reader.
• context-engineering — Starting a session, when output quality degrades, or when configuring rules files and project context.
• using-agent-skills — Deciding which skill or workflow applies, and how a set of skills should be discovered and invoked.
• handoff — Compacting a long conversation into a handoff document for a fresh agent session.
• chief-of-staff — Pursuing a long-running goal by coordinating sub-agents from a single session.
• team-composition-patterns — Deciding how many agents to use and which roles for a task.
• task-coordination-strategies — Breaking work into tasks with dependencies and clear descriptions for several agents.
• doubt-driven-development — Cross-examining a plan or an instruction file with a fresh-context adversarial review before it stands.
• avoid-ai-writing — Cleaning up instruction text and skill descriptions so they read plainly and precisely.

Commands

Invoke by name (for example /audit-instructions).

• audit-instructions — Check an instruction file against the repository and list stale, contradictory or unverifiable lines.
• new-skill — Interview the user, then draft a focused skill with a precise trigger description.
• context-budget — Estimate what the always-loaded instructions cost in context and propose cuts.

Sub-agents

• instruction-auditor — Delegate to verify claims in AGENTS.md or CLAUDE.md against the real repository and report drift.
• skill-reviewer — Delegate to review a draft skill for trigger quality, scope, length and clarity.

Scoped rules

• skill-files — Applies when creating or editing a SKILL.md file.
• instruction-files — Applies when editing AGENTS.md or CLAUDE.md instruction files.
• subagent-definitions — Applies when writing sub-agent definition files.

Where each agent finds it

Agent	Instructions	Skills and commands	Rules	Sub-agents
Claude Code	CLAUDE.md (imports AGENTS.md)	.claude/skills/	.claude/rules/	.claude/agents/
OpenAI Codex	AGENTS.md	.agents/skills/	listed in AGENTS.md	.codex/agents/
Cursor	AGENTS.md	.agents/skills/, .claude/skills/	.cursor/rules/	.claude/agents/
Google Antigravity	AGENTS.md	.agents/skills/	.agents/rules/	.agents/agents/
Gemini CLI	AGENTS.md via .gemini/settings.json	.agents/skills/, .gemini/commands/	listed in AGENTS.md	.gemini/agents/
GitHub Copilot	AGENTS.md	.agents/skills/, .github/prompts/	.github/instructions/	.claude/agents/
Windsurf (Devin)	AGENTS.md	.agents/skills/, .windsurf/workflows/	.windsurf/rules/	—
opencode	AGENTS.md	.agents/skills/, .opencode/commands/	listed in AGENTS.md	.opencode/agents/
Aider	AGENTS.md via .aider.conf.yml	—	—	—

Skills are stored twice (.agents/skills/ and .claude/skills/) because Claude Code reads only its own folder. Agents that read both may list a skill twice; delete the copy you do not need.

Credits

The skills come from these open-source repositories, each under its own licence. Their licence and a notice are kept beside every copied skill.

Skill	Source	Licence
writing-for-agents	mattpocock/skills at 49dd158	MIT
context-engineering	addyosmani/agent-skills at aeac41d	MIT
using-agent-skills	addyosmani/agent-skills at aeac41d	MIT
handoff	mattpocock/skills at 49dd158	MIT
chief-of-staff	mattpocock/skills at 49dd158	MIT
team-composition-patterns	wshobson/agents at 46891e7	MIT
task-coordination-strategies	wshobson/agents at 46891e7	MIT
doubt-driven-development	addyosmani/agent-skills at aeac41d	MIT
avoid-ai-writing	wshobson/agents at 46891e7	MIT

Everything else in this repository (instructions, rules, commands, sub-agents) is by Kitloom under the MIT licence.

Use at your own risk

Read every file before you use it. Agent instructions can make an agent run commands on your machine. This kit was checked with Kitloom's pre-download check, which looks for warning signs but cannot prove anything is safe.