Claude Code setup¶

How my Claude Code is set up. The two plugins that change day-to-day behaviour the most are caveman and ponytail. They complement each other:

• Caveman shrinks what the agent says.
• Ponytail shrinks what the agent builds.

Caveman doesn't touch code, and ponytail doesn't touch prose, so they don't overlap. Together you get terse explanations of minimal code.

Caveman¶

Repo: JuliusBrussee/caveman

Makes Claude drop filler, pleasantries and hedging and reply in short fragments ("Bug in auth middleware. Token expiry check use < not <=. Fix:"). Code blocks, commands, error messages and commit messages stay exact and are written normally. The README reports about 65% fewer output tokens on average across its benchmarks. Thinking tokens aren't affected, so the main win is replies that are faster to read.

Install¶

Run each command as a separate prompt inside Claude Code:

/plugin marketplace add JuliusBrussee/caveman

/plugin install caveman@caveman

Or from a shell:

claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman

On Claude Code it turns itself on at the start of every session through a hook, so you don't need to type /caveman.

Usage¶

It also ships cavecrew-* subagents (investigator / builder / reviewer) whose output comes back compressed, so the main context lasts longer.

Ponytail¶

Repo: DietrichGebert/ponytail

Puts a "lazy senior developer" inside the agent: someone who replaces fifty lines with one. Ask for a date picker and instead of installing a library and writing a wrapper component, you get:

<!-- ponytail: browser has one -->
<input type="date">

The ladder¶

Once the agent has read and understood the code involved, it stops at the first rung that works:

1. Does this need to exist?   → no: skip it (YAGNI)
2. Already in this codebase?  → reuse it, don't rewrite
3. Stdlib does it?            → use it
4. Native platform feature?   → use it
5. Installed dependency?      → use it
6. One line?                  → one line
7. Only then: the minimum that works

It never cuts input validation, error handling that prevents data loss, security, or accessibility. Deliberate shortcuts get a ponytail: comment that names their limit and how to upgrade later. That way they can be tracked instead of forgotten.

Install¶

Two separate prompts inside Claude Code:

/plugin marketplace add DietrichGebert/ponytail

/plugin install ponytail@ponytail

It needs node on the PATH of non-interactive shells for its hooks. Without it, the skills still work, but the always-on activation stays off.

Usage¶

Other plugins¶

From the official claude-plugins-official marketplace, which is available by default:

Plus obsidian-cli for reading and writing my Obsidian vault.

Skills¶

Skills are folders with a SKILL.md that Claude loads when a task matches its description. They live in ~/.claude/skills/, one folder per skill.

My own, written for research work:

They live in my dotfiles, so chezmoi update puts them on every machine where I run Claude Code. Skill scripts declare their Python dependencies inline (PEP 723) and run with uv run, so nothing has to be installed globally on each machine.

Installed with the skills CLI, which puts them in ~/.agents/skills/ and symlinks them into each agent's skills folder, so other coding agents can use them too:

npx skills add vercel-labs/skills      # find-skills: search for and install more skills
npx skills add ogulcancelik/herdr      # herdr: lets Claude drive Herdr panes and agents

A CLAUDE.md per project¶

Projects I work on with Claude have a CLAUDE.md at their root. Claude reads it at the start of each session. Keep it to what Claude can't work out from the code:

• the commands to build, run and deploy;
• the layout and why it's that way;
• the gotchas that already cost you time once.

For example, the CLAUDE.md for my homelab Docker stack says to always restart services from the repo root and never with -f compose/<file>.yml, because the included files can't see the root networks. The claude-md-management plugin above helps keep these files current.

Herdr integration¶

On every machine where I run Claude inside Herdr, herdr integration install claude adds a SessionStart hook, so Herdr knows which Claude session lives in which pane. See Remote machines.

Safety hook: confirm recursive rm¶

A PreToolUse hook stops any Bash command containing a recursive rm and asks me on the terminal before it runs. Claude can't approve it on my behalf.

#!/usr/bin/env bash
# Hook: intercept rm -r / rm -rf commands and require explicit user confirmation.
# Receives tool input JSON on stdin; exits 0 to allow, exits 2 to block.

set -euo pipefail

input=$(cat)
command=$(echo "$input" | python3 -c "import sys,json; d=json.load(sys.stdin); print(d.get('command',''))" 2>/dev/null || true)

# Match rm with recursive flags anywhere in the command
if echo "$command" | grep -qE '(^|[;&|`(]|\bsudo\b)\s*rm\s+[^;]*-[a-zA-Z]*r'; then
  # Write prompt directly to the terminal, bypassing stdout (which Claude reads)
  exec 3>/dev/tty
  echo "" >&3
  echo "⚠️  Claude wants to run a recursive rm command:" >&3
  echo "   $command" >&3
  echo "" >&3
  printf "Allow this? [y/N] " >&3
  exec 3>&-

  # Read the answer from the terminal
  read -r answer </dev/tty

  if [[ "$answer" =~ ^[Yy]$ ]]; then
    exit 0
  else
    echo "User denied recursive rm. Command was: $command"
    exit 2
  fi
fi

exit 0

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "bash ~/.claude/hooks/rm_confirmation.sh",
            "timeout": 30
          }
        ]
      }
    ]
  }
}

Statusline¶

The statusline shows the caveman and ponytail mode badges plus my plan usage on one line:

[CAVEMAN] [PONYTAIL]  ·  session 42% · 2h10m · weekly 18% · $3.2 session

• Caveman badge: [CAVEMAN], or [CAVEMAN:ULTRA] etc. at levels other than full. If /caveman-stats has run at least once, the lifetime tokens saved follow the badge. Set CAVEMAN_STATUSLINE_SAVINGS=0 to hide that.
• Ponytail badge: [PONYTAIL] in green, or [PONYTAIL:LITE] etc. at other levels. ULTRA shows in amber. It comes from hooks/ponytail-statusline.sh, which ships with the plugin. It doesn't use the JSON from Claude Code, so the wrapper calls it without piping anything in.
• Usage: how much of the 5-hour session quota is used and when it resets, the weekly quota, and this session's cost. Percentages turn yellow at 75% and red at 90%.

Claude Code only accepts one statusLine command, so a small wrapper script builds the line. It's the same on macOS and Linux. The usage numbers come straight from the JSON that Claude Code sends to the statusline command on stdin (rate_limits and cost), and jq formats them. jq is the only extra dependency: it's preinstalled on recent macOS, and on Linux it's apt install jq or equivalent.

#!/bin/bash
# Combined statusline: caveman + ponytail badges + usage from Claude Code's own rate_limits JSON.
INPUT=$(cat)

CAVEMAN=$(printf '%s' "$INPUT" | bash "$HOME/.claude/plugins/marketplaces/caveman/src/hooks/caveman-statusline.sh" 2>/dev/null)
PONYTAIL=$(bash "$HOME/.claude/plugins/marketplaces/ponytail/hooks/ponytail-statusline.sh" 2>/dev/null)
BADGES="$CAVEMAN${CAVEMAN:+${PONYTAIL:+ }}$PONYTAIL"

USAGE=$(printf '%s' "$INPUT" | jq -r '
  def col(p): if p >= 90 then "\u001b[31m" elif p >= 75 then "\u001b[33m" else "\u001b[32m" end;
  def eta(t): (t - now | floor) as $s
    | if $s < 60 then "\($s)s" elif $s < 3600 then "\($s/60|floor)m"
      elif $s < 86400 then "\($s/3600|floor)h\(($s%3600)/60|floor)m" else "\($s/86400|floor)d" end;
  [ (.rate_limits.five_hour // empty | col(.used_percentage) + "session \(.used_percentage|round)%\u001b[0m\u001b[2m · \(eta(.resets_at))\u001b[0m"),
    (.rate_limits.seven_day // empty | col(.used_percentage) + "weekly \(.used_percentage|round)%\u001b[0m"),
    (.cost.total_cost_usd // empty | "\u001b[2m$\(. * 100 | round / 100) session\u001b[0m")
  ] | join(" · ")' 2>/dev/null)

if [ -n "$BADGES" ] && [ -n "$USAGE" ]; then
  printf '%s  \033[2m·\033[0m  %s' "$BADGES" "$USAGE"
elif [ -n "$BADGES" ]; then
  printf '%s' "$BADGES"
else
  printf '%s' "$USAGE"
fi

Make the wrapper executable with chmod +x ~/.claude/statusline-combined.sh, then point Claude Code at it:

{
  "statusLine": {
    "type": "command",
    "command": "~/.claude/statusline-combined.sh"
  }
}