setup-git-lite - SKILL.md Agent Skill

name: setup-git-lite description: 'Disable Claude Code built-in git instructions and inject a curated 280-tok minimum via SessionStart hook. Saves ~2,200 tok/session + ~1,700 tok/call.' when_to_use: | AI may auto-invoke for SAFE subcommands only — dismiss-banner, undismiss-banner, status, help. AI MUST NOT auto-invoke DESTRUCTIVE subcommands — install, revert. Destructive subcommands require the user to explicitly type the slash command.

Manage Claude Code's built-in includeGitInstructions token consumption.

Subcommand routing

Parse the first argument from $ARGUMENTS:

Subcommand	Risk	AI-allowed	Behavior
`install`	🔴 destructive	❌ user-only	Modifies ~/.claude/settings.json + shell profile. Requires explicit `<command-name>claude-code-token-saver:setup-git-lite</command-name>` tag with `<command-args>install</command-args>`
`revert`	🔴 destructive	❌ user-only	Aggressive cleanup. Same gating as install
`status`	🟢 safe	✅ allowed	Read-only diagnostic
`dismiss-banner`	🟢 safe	✅ allowed	Suppresses the recommendation banner (preferences.json flag)
`undismiss-banner`	🟢 safe	✅ allowed	Re-enables the banner
`help`	🟢 safe	✅ allowed	Prints usage
(empty)	🟢 safe	✅ allowed	Run `status`, then append one-line usage hint: "Subcommands: install, revert, status, dismiss-banner, undismiss-banner, help"

Invocation guard — enforce for destructive subcommands only

If the subcommand is install or revert:

Verify explicit invocation: the triggering user message MUST contain <command-name>claude-code-token-saver:setup-git-lite</command-name> along with <command-args>install</command-args> (or revert). If these tags are absent (i.e., you inferred the intent from natural-language context), STOP and respond:

"This action modifies your global Claude Code settings and shell profile. Please invoke explicitly by typing /setup-git-lite install (or revert)."
User confirmation before execution: even when the tag is present, print a one-line summary and ask for a literal "yes":
"Install will:
- Set includeGitInstructions: false in ~/.claude/settings.json
- Append a marker block to your shell profile exporting CLAUDE_CODE_DISABLE_GIT_INSTRUCTIONS=1 Proceed? (type yes to confirm)"
Only proceed on a literal yes. Any other response — cancel.

Safe subcommands (status, dismiss-banner, undismiss-banner, help) skip the guard entirely.

Execution

After (if required) passing the guard + confirmation, run:

node ${CLAUDE_PLUGIN_ROOT}/scripts/setup-git-lite.js <subcommand>

Relay stdout verbatim to the user. On non-zero exit, surface stderr.

Language

Detect the user's language from the conversation. Translate ALL user-facing output into that language. The script outputs English — you MUST translate before relaying. Keep technical identifiers verbatim (file paths, env var names, command names, setting keys).

Output conventions

status: translate stdout into the user's language, then relay.
install: translate stdout, then add a short line pointing to README for the rationale.
revert: translate stdout, then note "Your current shell env may still have the variable set — restart the shell or run unset CLAUDE_CODE_DISABLE_GIT_INSTRUCTIONS." (in the user's language).
dismiss/undismiss: translate stdout, then relay.
help: translate stdout, then relay.

Error handling

Exit code 1 (general error): print stderr, suggest running status.
Exit code 2 (unknown subcommand): surface the hint from the script, no further work.

Design reference

The hook content and behavior are documented in the claude-code-token-saver README (Git instructions section). Key points for the user if they ask:

Original CC built-in injects ~2,200 tokens per session (git status snapshot + full commit/PR workflow).
Our replacement is ~280 tokens (11 override rules + compact git state line with file list).
88% token reduction, 95%+ of the safety rules preserved.
Other style/workflow guidance is dropped because Claude's training already covers it (e.g., PR title conventions, HEREDOC format awareness, gh usage).