Claude Code tips from social

Comprehensive Guide to Claude Code Optimization and Workflows

Claude Code is a powerful terminal-based tool that can be optimized using custom plugins, advanced commands, and behavioral settings. Effective configuration allows developers to manage token limits, automate repetitive workflows, and delegate tasks to autonomous subagents [1][19][20].

Core Configuration and Project Files

Project-level memory and rules are essential for maintaining consistent behavior across sessions. Creating a permanent project memory file via the `/init` command or by writing a `claude.md` file establishes an always-on table of contents and a set of persistent rules [2][3][5][19]. Structure rules using path-specific paths and HTML comments to minimize token usage and load instructions dynamically [15]. Additionally, filling out personal preferences in settings helps define user identity and output styles [8].

Token and Context Management

Performance and output quality drop as the context window fills up, with some sources noting performance degradation starting around 30% capacity [2][16][19]. Several built-in terminal commands and plugins help manage context limits: - Run `/compact` or clear sessions often when approaching context limits [2][5][16]. - Use `/context` to check remaining space and what is currently in the context window [5]. - Use the `/rewind` command (or `/re`) to jump back to a previous checkpoint if Claude messes up or outputs break, avoiding context bloat [5][15][20]. - Utilize plugins like Ponytail, which cuts token usage by over 50% without losing accuracy, and Headroom, which compresses and removes unnecessary data between the user and the model [1][9]. - Open-source or alternative commands like `/caveman` can cut about 75% of response tokens by eliminating fluff [10].

Commands and Behavioral Control

Switching output styles and behavioral commands alters how Claude reasons and responds. Use option-tab to switch modes, and add `/ultrathink` or `/deepthink` to allow Claude to reason up to 10 times deeper before answering complex prompts [2][10]. The `/ghost` command removes AI tells, preambles, conclusions, and m dashes, while `/output-style` (or switching to explanatory or concise output styles) helps users understand coding decisions or speed up workflows [5][10][15]. To automatically save session learnings, users can type `/reflect` at the end of a session, a prompt habit that can be saved as a skill to update memory and workflow preferences [4].

Skills, Subagents, and Automation

Skills act as step-by-step playbooks or reusable markdown workflows that load only when triggered, and can be found on platforms like `skills.sh` [3][19][20]. Skills can be locked using `disable model invocation` to require manual execution, or created by talking to Claude or recording workflows [6][20]. Subagents handle side tasks in separate windows and report back with answers, while custom subagents allow Claude to delegate tasks to specialized markdown-defined agents [3][20]. Hooks let developers trigger bash commands and Python scripts during specific lifecycle events like pre-tool and post-tool use [20]. For autonomous building, the `goal` command runs continuously until a condition is met, whereas the `loop` command runs on a fixed timer [14].

Integrations, Connectors, and MCP Servers

Claude Code connects to external data sources, applications, and tools through integrations and the Model Context Protocol (MCP). Users can add connectors for meeting notes, email, calendar, Slack, Notion, and finance software in settings [6][8]. For MCP servers, developers can use commands such as `Cloud MCP add` (with a transport type and name) and `Cloud MCP list` to manage servers, though newly added MCP tools require a Claude Code session restart [11]. While Claude Code can call existing REST APIs directly, MCP standardizes how applications discover and use capabilities, keeping secrets like API keys inside the MCP server so the LLM never sees them [12][13]. Tools like Arcade act as a runtime for MCP to handle per-user authentication and authorization [12].

Advanced Execution and Workflows

Advanced development workflows incorporate parallel environments, Git worktrees, and headless execution. Git worktrees enable running multiple isolated Claude sessions in parallel across separate copies of a repository without interference [20]. Headless mode allows running Claude non-interactively within bash scripts and piping outputs into JSON format using tools like jq [20]. Team members also utilize Claude Tag within Slack for broad product context, and deploy software factory workflows where an adversarial review agent verifies work done by a build subagent [14][18].

What is not covered

Detailed pricing tiers or subscription models beyond the $20 monthly pro plan are largely omitted [6]. Specific technical documentation for every third-party CLI tool mentioned (such as Supabase, Vercel, or Warp) is not fully detailed [14].