The agent toolkit

A portable set of skills, slash commands, safety hooks, and cross-agent rules that you install for any coding agent, either globally (every repo on your machine) or into one specific repository. Once installed, the agent already knows how the owner thinks and works, already knows your reusable systems (error-alert design, RLS security, query/page performance, the money-truth doctrine, SEO, the review standards, and more), and follows the same rules (no accidental commits/pushes/tree-destroys, no em dashes, verify a DB write against the live schema first, decide instead of asking when the answer is derivable).

The point of the toolkit: nothing learned in one project has to be rewritten in the next. Every lesson that is not tied to a single product lives here, so a new project starts with an agent that already knows the owner and the way he builds.

It works for Claude Code, Cursor, and Codex (and any other agent that reads an AGENTS.md). Skills use the same SKILL.md format all three understand; the installer just puts them in each agent's own skills folder.

This folder IS the toolkit. install.py copies it where each agent looks.

---

What is in here

Folder / file	What it is
skills/	The skill library (one folder per skill, each with a SKILL.md). Portable across Claude Code, Cursor, and Codex.
commands/	Slash commands (e.g. /problem-cause-solution). Claude Code.
hooks/	PreToolUse + Stop hooks (git-safety gate, no-em-dash, code-quality/security checks). Claude Code only. Self-contained, no global setup needed.
settings.json	Claude Code settings that wire the hooks and set permissions. The installer MERGES this, it does not clobber.
AGENTS.md	The owner's standing rules (the ones the hooks enforce, plus his working agreements) and the skill index by category, written as instructions. Cursor and Codex read it in place of the hooks.
CLAUDE.template.md	The starter CLAUDE.md for a new repository: the owner's standing rules, the database gate, the git rules, a which-skill-to-read-when table, and a project-facts section for the first agent to fill in. The installer writes it to the repo root as CLAUDE.md when the repo has none, and never overwrites an existing one.
install.py	The installer. Handles all agents and both scopes (global / one repo).
README.md	This file.

What the skill library covers

AGENTS.md holds the full index by category. In short:

Area	What is in it
The owner	How he thinks, decides, writes, and what he counts as done (how-the-founder-thinks, founders-style, working-with-the-owner, working-agreements). Read these first in a new project.
Standards library	About 40 standard-* skills (security, config and secrets, design system, accessibility, API and database design, testing, CI/CD, backups, privacy, mobile apps, and more). standards-index lists them all.
Engineering discipline	Performance, query efficiency, RLS, auth, URL structure, timezones, verifying before calling something done, safe database writes, debugging.
Money systems	The money-truth doctrine, POS checkout, payment processor integration, SaaS subscriptions, commissions, memberships and packages, reports.
AI agents	Internal write-agents, the model/feature platform, multi-agent assistants, AI data migration.
Phone, voice, messaging	Retell and self-hosted voice agents, voice prompting, number provisioning, WhatsApp (managed and self-hosted), proxies, email.
Running a product	Support triage, help center upkeep, onboarding, public API upkeep, ticketing, website builder, release emails.
Deploy and devices	Release sequence, deploy freshness, maintenance banner, alert patrol, iOS from Windows, Android, Tauri desktop, GPU runs.
Design, motion, mobile	UI/UX, Apple design review, the taste skills, motion and animation, mobile view optimization, product animation, video.
Growth	SEO, local SEO, the Google Ads set.
Third-party packs	The Stripe skills and a planning/debugging/TDD process pack.

Where the skills came from. Many were extracted from skills and rule files written inside specific projects, then rewritten to drop product names, paths, and identifiers while keeping the lessons. When you learn something reusable in a project, add it here (see the skill-creator skill) instead of leaving it in that project's own skills folder.

Never put secrets or account identifiers in a skill. Refer to them by environment variable name or file path only.

The hooks require Python on PATH as py (Windows) or adjust to python3 (macOS/Linux). See "Non-Windows" below. install.py runs under either.

---

Install it

There are two commands: install globally (applies to every repo on this machine) or install into one specific repository. Each works for all agents at once, or a single agent you name.

By default it installs for Claude Code only. Add --agent to include Cursor and/or Codex.

1. Install globally (every repo on this machine)

py /path/to/toolkit/install.py --global

This installs into ~/.claude/ (Claude Code). It does not touch your other agents unless you ask for them.

2. Install into a specific repository

# run from inside the repo you want to set up:
py /path/to/toolkit/install.py .

# or name the repo:
py /path/to/toolkit/install.py C:/path/to/repo

This installs into that repo's .claude/ (Claude Code).

Include Cursor and/or Codex (opt in)

Add --agent. Skills then also land in that agent's skills folder, plus an AGENTS.md (at the repo root, or your home dir for --global) carrying the rules.

py install.py --global --agent all             # all three agents, globally
py install.py . --agent cursor                 # add Cursor to this repo
py install.py . --agent claude,cursor,codex    # all three, this repo

Valid agents: claude, cursor, codex, all. Default: claude only.

Or just tell the agent

In a repo that already has the toolkit (or that can reach a clone of it), say:

install the toolkit

The agent runs the install-toolkit skill, which calls install.py.

Then

Restart the agent so it picks up the new skills (and, for Claude Code, hooks).

---

What the installer does per agent

Agent	Skills go to	Also gets
Claude Code	.claude/skills/	.claude/commands/, .claude/hooks/, a MERGED .claude/settings.json that wires the hooks, and (per-repo installs) a starter CLAUDE.md at the repo root if there is none
Cursor	.cursor/skills-cursor/	AGENTS.md at the root (the rules the hooks enforce, as instructions)
Codex	.codex/skills/	AGENTS.md at the root

• Merges, never clobbers. settings.json unions permissions.allow/deny/ additionalDirectories and adds the hooks, keeping everything the target had. It backs up the old settings.json to settings.json.bak-<timestamp> first.
• Never overwrites an existing AGENTS.md (if the repo already has one, it is left alone; the rules also live in the skills).
• Is idempotent. Safe to re-run to update: it refreshes files and de-dupes, so it never doubles a hook or a permission.
• Only writes under each agent's config dir (.claude / .cursor / .codex) and the root AGENTS.md. It never deletes skills the target added itself.

Why Cursor and Codex get AGENTS.md instead of hooks: Claude Code's hooks are Python scripts Claude runs before a tool call. Cursor and Codex do not run them. So the toolkit gives those agents the same intent (no em dashes, do not commit/push on your own, verify DB writes, do not swallow errors) as written rules in AGENTS.md, which both agents read automatically.

---

Keeping the toolkit somewhere your team can pull it (recommended)

This toolkit lives at https://github.com/Automate-Anything/claude-toolkit (private). Any developer with access can install or update it in any project:

# one-time: clone the toolkit somewhere
git clone https://github.com/Automate-Anything/claude-toolkit C:/dev/claude-toolkit

# in any project, install/update from that clone:
py C:/dev/claude-toolkit/install.py .

To update everyone later: push changes to the toolkit repo, and each dev re-runs git pull in their clone plus py <clone>/install.py . in their projects (or just says "update the toolkit" to the agent). Because the merge is idempotent, re-installing only refreshes and never duplicates.

If you prefer no git, each developer keeps a local copy at a known path (e.g. c:\dev\.claude) and installs from there; they just have to update that copy themselves.

---

After installing: adapt the "design pattern" skills

Some skills describe a system you should BUILD, not one that already exists (error-alert-system, security-rls-permissions, money-truth-doctrine, and the "how it should be wired" sections of others). They open with an "adapt before use" note. Once you build the real system in a repo, update that skill to name the actual files/tables/functions so the next agent finds real code, not just the pattern.

Optional: turn on the database-write gate

hooks/pretooluse_bash.py can require an agent to re-attest it verified the live schema before running a DB write. It is off by default. To enable it, add your repo's SQL-runner script basename to DB_RUNNER_SCRIPTS at the top of that file, e.g. DB_RUNNER_SCRIPTS = ("run_sql.py",).

---

Non-Windows machines

The Claude Code hooks are invoked as py "..." in settings.json. On macOS/Linux py may not exist. Two options:

1. Create a py shim: alias py=python3 (or a small wrapper on PATH), or
2. After installing, edit .claude/settings.json and replace py with python3 in each hook command.

install.py itself runs under either py or python3. (Cursor/Codex do not run the hooks, so this only affects Claude Code.)

---

Uninstall

Everything the installer wrote lives under the agent config dirs (.claude / .cursor / .codex) and the root AGENTS.md. To remove: delete the toolkit's skills from each agent's skills folder, delete .claude/hooks/, restore .claude/settings.json from its settings.json.bak-<timestamp>, and delete AGENTS.md if the toolkit created it. All of it is local and reversible.