Skip to content

Repository files navigation

claudzilla

Claude Code, tamed: terse answers, small diffs, verified work, guarded git.

One command turns a stock ~/.claude into a careful senior dev's setup.

Release Tests License: MIT Platforms

Install

curl -fsSL https://raw.githubusercontent.com/HT0710/claudzilla/main/install.sh | bash

Requires: git, curl, tar, perl and Claude Code. Missing node or rtk is installed to ~/.local without sudo; keep ~/.local/bin on your PATH. Native Windows is not supported; use WSL.

Pick parts

On a fresh install it asks Full (recommended) or Custom; re-runs keep your saved choice. Full installs every part; Custom lets you untick parts. Run by Claude (no terminal), it stops and prints the questions for Claude to ask you. A fresh --add hud installs only hud plus the updater (no rules, no core settings).

flag does
--full install every part
--custom pick parts from the menu
--add a,b add parts to your install
--remove a,b remove parts from your install
--uninstall remove claudzilla (asks first)
--yes skip the --uninstall confirmation

Flags go after bash -s --. For example, only the status line:

curl -fsSL https://raw.githubusercontent.com/HT0710/claudzilla/main/install.sh | bash -s -- --add hud
part covers
core rules, hooks, settings, pr and rule-review skills
hud statusline, themes, MesloLGS NF font
rtk token-saving command filter
caveman terse replies
ponytail smallest-change coding
dividers response dividers
panel side panel
picks suggested next steps
toast update and usage notices
attribution commit credit

core is not in the menu; --add and --remove accept it. The updater always ships; it also carries the guide, /claudzilla report and /claudzilla compact. Your choice is saved, see How install works.

Safe to re-run. Anything replaced is backed up to ~/.claude/.claudzilla-backup/.

New here? Read GUIDE.md, or type /claudzilla in Claude.

What you get

Answer first, evidence cited, one fix proposed, nothing edited until you say go, and "fixed" only after the test ran.

Every picture is a real capture of one small demo repo, shot with permission prompts off (--dangerously-skip-permissions, hence the red bypass permissions on line); only the account email is blanked out.

Answers you can scan

  • Terse replies: TL;DR first, tables over prose, every claim backed by a file:line or a command, or marked unverified.

    Claude Code with claudzilla answering why a test fails: TL;DR first, then a cause table citing file:line and the real test output
  • Investigate, then ask: root cause and one recommendation before any multi-file edit.

    A decision table with options A and B, a bold recommendation with its source, and a Next line
  • Picks card: a reply that ends in a decision shows a card above the prompt; press a digit to answer (details).

    The picks card above the prompt; pressing 1 puts option A and its text in the prompt box, then Enter sends it

    Real speed. Press 1: the option's id and text go into the prompt box; Enter sends it. Needs the picks part.

  • Dividers and file links: lines frame the TL;DR, Next: and each ---, fitted to the chat width. A file path in a reply is a link: click it to see those lines in the panel (details).

    Clicking auth/token.js:3 in a reply opens the panel's File tab with line 3 marked

    Real speed. Click auth/token.js:3: the File tab shows the lines around it, line 3 marked ▶. Needs the fullscreen layout and the dividers and panel parts. Your terminal's own link click, such as ctrl+click, opens the file; Claude Code does not.

Code that stays small

  • Lazy by default: the smallest change that fully works (ponytail).

    A one-line test fix shown as a +2 -2 diff, then Claude reports the test now passes
  • Disciplined workflow: debugging, TDD, plans and reviews via superpowers.

  • Fewer tokens: short-form voice (caveman), and every Bash call is filtered through rtk.

Git that can't bite

  • Guard hook: blocks force-pushes, discarding work and commits on main.

    Asked to commit straight to master, Claude reports the guard hook blocked it and names the opt-in setting
  • Verify before push: push and PR wait until verification ran; "done" claims with no check run after the last edit get sent back.

    The stop hook sends a done claim back; Claude reruns the checks, including one that proves the test fails if the code is broken again
  • Clean history: Conventional Commits enforced, Claude session links stripped from commits and PRs (details).

A better terminal

  • Statusline: cwd, branch, context / 5h / weekly meters, model, effort, active skill, claudzilla version.

    The status line during one turn: the context meter climbs and the loaded skill appears beside the model

    3× speed. One real turn: the context meter climbs and the skill Claude loaded appears. The 5h and wk meters show only on a Claude subscription login, such as Pro, Max or Team. The symbols need the MesloLGS NF font; without it, basic symbols.

  • Prompt panel: /panel shows what Claude plans and does, lists this chat's prompts, previews files and shows git; click to jump, reuse or rewind (details).

    The panel's Overview tab during a fix: planning, then ASK, PLAN and steps that tick off one by one until done

    2× speed. After planning…, Claude's reading of the request (ASK, PLAN) fills in, and each step goes ○ → ● → ✓ until the header reads ✓ done. Docks beside the chat in the fullscreen layout, from 110 columns. Costs: a short tool call per report (the panel's intent option in /config turns it off), and while the Overview tab shows, an Opus request at low effort after each prompt to sort topics.

  • Toasts: updates install in the background and say so; a toast also shows when the 5-hour or weekly limit passes 50%, 75% and 90%, and when context is worth compacting (details).

  • Theme + settings: muted dark truecolor theme, opus[1m], normal permission prompts.

  • Local trace: a record of what claudzilla's parts did, kept in ~/.claude/claudzilla-trace/, never uploaded; {"rulesGuard":{"trace":false}} in ~/.claude/claudzilla.json turns it off. See the guide topic bugs (/claudzilla bugs).

More screenshots

Panel Git tab: the branch, changed files and recent commits.

The panel's Git tab: master with one changed file under CHANGES and one commit under COMMITS

Panel History tab: this chat's prompts, newest first, with rewind and latest.

The panel's History tab listing two prompts under Today

Guide pane: /claudzilla or alt+g, topics beside a section.

The guide pane: topic list grouped by part, the What claudzilla is section under it

Skills and commands

command does
/claudzilla [topic] the guide, one topic at a time
/claudzilla report [what happened] drafts a problem report from the local trace (nothing uploaded); files an issue only on your yes
/claudzilla compact drafts a /compact line into the prompt box
/panel side panel: Claude's plan and steps, this chat's prompts, files, git
/rule-review how often each rule slipped; weekly summary at startup; on your yes, shares the counts (no text) as a claudzilla issue
pr skill pre-PR checks + description template, runs before every PR

Customize

file for
~/.claude/CLAUDE.local.md instructions for this machine only
~/.claude/settings.overrides.json settings for this machine only, merged last
~/.claude/claudzilla.json turn guard checks on or off

Full reference, install internals and keybindings: CONFIGURATION.md.

Update

Each push to main with a feature, fix or refactor ships a release, once the tests pass. At most once a day, a new session or /clear installs the newest release in the background, unless the clone has local changes; then you get a notice. After an update, /claudzilla news lists what changed. To update by hand (latest main, may be ahead of the release):

cd ~/claudzilla && git pull && ./install.sh

Uninstall

~/claudzilla/install.sh --uninstall

Removes claudzilla's links, settings entries, keys, plugins and record from ~/.claude, and undoes what the install recorded (node, fonts, shell lines). It lists what is still there (rtk, copies it did not install, the clone) with the commands to delete them. Claude runs it with --yes after you confirm.

Credits

Built on caveman, ponytail, superpowers and rtk.

License

MIT

Releases

Packages

Contributors

Languages