Files
Claude-Code-Game-Studios/.claude/docs/setup-requirements.md
T
DonchitosandClaude Opus 5.5 be8993bbc5 Release 1.1.3 — long skills keep their later steps
Fix release for 1.1.2: the 14 longest skills keep their later phases
after a compaction, a question nobody answered is never taken as an
answer, settings read the same from a subfolder, the commit check works
for a project in a subfolder of a larger repository, every skill opens
with a title, and the shipped allow rules no longer approve deleting a
branch or writing files without asking. Details in CHANGELOG.md; upgrade
steps in UPGRADING.md (v1.1.2 → v1.1.3).

Refs #83
Refs #133

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0185en4t76fPn7315q4c3nko
2026-10-08 17:52:48 +11:00

6.3 KiB

Setup Requirements

This template requires a few tools to be installed for full functionality. Hooks carry on without an optional tool and lose only the check it powers. The required ones are required: without Python 3 every setting falls back to its default and a commit that stages data JSON is blocked, and without Bash no hook runs at all.

Required

Tool Purpose Install
Git Version control, branch management git-scm.com
Claude Code AI agent CLI Native installer: curl -fsSL https://claude.ai/install.sh | bash (macOS, Linux, WSL) or irm https://claude.ai/install.ps1 | iex (Windows PowerShell). npm also works: npm install -g @anthropic-ai/claude-code (Node.js 22+). See code.claude.com/docs/en/setup
Python 3 Reads project.yaml for every skill and hook python.org. Without it every setting silently falls back to its default, and a commit that stages data JSON is blocked because it cannot be validated
Bash Runs every hook Built into macOS and Linux. On Windows, install Git for Windows — Claude Code runs without it, but no CCGS hook does
Tool Used By Purpose Install
jq Hooks (8 of 12) JSON parsing in commit/push/asset/agent hooks See below

Installing jq

Windows (any of these):

winget install jqlang.jq
choco install jq
scoop install jq

macOS:

brew install jq

Linux:

sudo apt install jq     # Debian/Ubuntu
sudo dnf install jq     # Fedora
sudo pacman -S jq       # Arch

Platform Notes

Windows

  • Git for Windows includes Git Bash, which provides the bash command used by all hooks in settings.json
  • If Claude Code can't find Git Bash, set CLAUDE_CODE_GIT_BASH_PATH to its bash.exe in the env block of your settings (for example C:\\Program Files\\Git\\bin\\bash.exe in JSON)
  • Hooks use bash .claude/hooks/[name].sh — this works on Windows because Claude Code invokes commands through a shell that can find bash.exe

macOS / Linux

  • Bash is available natively
  • Install jq via your package manager for full hook support

Verifying Your Setup

Run these commands to check prerequisites:

git --version          # Should show git version
bash --version         # Should show bash version
jq --version           # Should show jq version (optional)
python --version       # Required: Python 3 (python3 or py also work)

What Happens Without Optional Tools

Missing Tool Effect
jq Nothing is lost: every hook falls back to grep to read its JSON input. jq is only more robust with unusual input.
Python 3 Not optional (see Required). Every setting falls back to its default, because project.yaml is read through Python; the asset hook skips JSON validation and says so; a commit that stages data JSON is blocked.
Both As without Python.

Optional Performance Settings

CCGS deliberately ships no values for the settings below. Each one's best value depends on your machine, your Claude plan and how you work, so a value committed here would be a guess every user inherits without noticing. Set the ones you want in .claude/settings.local.json, which is gitignored and yours alone.

Setting What it does When to change it
promptCacheTtl How long the main conversation's prompt cache lives. Raise it if you work in long sessions with gaps; the cache surviving a break avoids re-sending context.
subagentPromptCacheTtl The same, for subagents and other off-conversation requests. Worth raising on a 49-agent project like this one, where team-* skills spawn repeatedly.
autoCompactWindow How full the context gets before Claude Code compacts it. Lower it if compaction keeps surprising you mid-task; raise it if you would rather compact less often and keep more history.
skillListingBudgetFraction How much context the skill listing may occupy. CCGS ships 74 skills, so the listing is not small. Lower it if you want more room for work; skillListingMaxDescChars trims each description instead.

sandbox.enabled isolates shell commands from your filesystem and network. It is worth turning on, but it runs on macOS, Linux and WSL2 only — it is not available on native Windows, which is CCGS's primary platform.

One setting CCGS does ship

permissions.defaultMode is set to default in .claude/settings.json, on purpose. A terminal session started at the project root then starts in Claude Code's ask-first mode, so file edits wait for your approval even if your own config sets acceptEdits or auto.

It sets the mode a session starts in. It is a backstop, not a guarantee, and these routes go around it:

  • Starting in a subfolder such as design/ or the code root (src/, Assets/ or Source/, depending on the engine). Claude Code loads .claude/settings.json only from the folder it starts in, so that session has none of this project's hooks, deny rules or permission mode. Start at the root.
  • Resuming a session. A resumed session keeps the mode it ended in.
  • Switching during a session: Shift+Tab, answering "Yes, and switch to auto mode", or allowing all edits for the session.
  • claude --permission-mode auto on the command line.
  • The VS Code extension, which never reads a project's .claude/settings.json for the mode a conversation starts in. Set claudeCode.initialPermissionMode to default in your VS Code user settings to get the same behaviour.
  • The Desktop app, which remembers the mode you last picked for a folder and uses that instead.

In any mode that accepts edits without asking, the framework's "May I write?" is an instruction to the model, not a stop.

To start terminal sessions in a different mode, set permissions.defaultMode in .claude/settings.local.json, which takes precedence. Terminal sessions accept any mode from it except auto and bypassPermissions. Understand what you are turning off first.

Claude Code works with any editor, but the template is optimized for:

  • VS Code with the Claude Code extension
  • Cursor (Claude Code compatible)
  • Terminal-based Claude Code CLI