Claude Code Power-Ups
I recently exported a week’s worth of messages from the #claude-code Discord channel, and asked Gemini 2.5 Pro to surface tips & tricks beyond Anthropic’s documentation.
Models#
- The default model on Max plans pairs Opus with Haiku for subtasks, potentially switching to Sonnet if Opus limits are near.
- Users on the $100 Max (5x) plan often hit Opus limits within 1-4 hours, frequently switching to Sonnet 4 (
/model sonnet-4-20250514
) or using Opus for planning/review and Sonnet for execution. - The $200 Max (20x) plan provides about 5 hours of intensive Opus usage before limits are reached.
- Pin specific model versions (e.g.,
claude-3-7-sonnet-20250219
) viaANTHROPIC_MODEL
. - Usage resets every 5 hours (local time); track with
/status
. - Use community scripts to analyze local session data (
~/.claude
) for cost insights.
Prompting#
- Encourage deeper thinking with keywords like “think harder” or “ultrathink.”
- Clearly document complex tasks in markdown (steps, subtasks, validation criteria).
- Explicitly instruct Claude: “Don’t make any code changes yet” to keep discussions theoretical.
- Workflow: Claude creates a plan → you refine → Claude executes.
Permissions#
- Bypass repetitive permission checks with
claude --dangerously-skip-permissions
(recommended in isolated Docker containers or VMs). - Deny specific permissions (
Bash(*:.env)
) to protect sensitive files.
IDE Integration#
- VS Code extension (via terminal or npm
.vsix
) provides live diffs and editor context. - JetBrains integration works well; external terminal recommended for stable output.
- WSL users should use Linux paths (
/home/...
) for better performance.
Terminal & Shell#
tmux
is recommended for efficiently managing multiple Claude sessions simultaneously.- Low-latency terminals like Alacritty, Ghostty, and Wezterm offer smoother interactions compared to standard options.
- Use
/terminal-setup
on macOS terminals (Terminal.app, iTerm2) to enable convenient shortcuts like Shift+Enter for new lines.
Session Management#
- Quickly resume sessions using
claude --resume
andclaude --continue
. - Regularly clean old session files (
~/.claude/projects/
) to improve startup performance. /clear
wipes current context but retains session resumability.
Community Tools#
claude-simone
: Structured, file-based project/task management.- claude-squad offers automated workflows for managing parallel agents efficiently, enhancing productivity by structuring tasks clearly.
- Cost-tracking scripts (Python, Node.js, Electron) provide insights from local session files.
- MCP tools:
isitdown
(website checks) andask-ollama.sh
(local Ollama task delegation).
Multi-Agent#
- Run concurrent Claude instances for separate roles (planning, execution, testing).
- Enable
parallelAgents
for managing concurrent subtasks efficiently. - Coordinate agents via shared documents like
PROGRESS.md
. - Sub-agents (
Task()
tool) provide isolated contexts for specific investigations.
Workflow#
- Adopt a structured cycle: Research → Plan → Review → Execute.
- Refine code iteratively with Claude critiques (e.g., persona-driven feedback like “as John Carmack”).
- Leverage external tools by instructing Claude to create scripts that interface with CLI tools (
gh
, Ollama).