Claude Code Digest — Apr 05–Apr 08

Claude Code's hidden /compact flag cuts token usage by 90% for lightning-fast iterations.

90% token usage reduction with Claude Code's /compact flag

Trending Now

✨ Kerf-CLI: Cost Tracking Made Easy
With Kerf-CLI, users gain precise control over Claude Code spending with a local SQLite dashboard, making it easier to enforce budgets and identify wasted spend. 🔥 Hazmat: Safe Autonomous Sessions on macOS
Hazmat makes the `--dangerously-skip-permissions` flag viable by adding OS-level security, allowing for fully autonomous Claude Code sessions without compromising safety. 📈 Opus+Codex Crossover Point
Switch from Opus to Codex at 600 lines of code to optimize costs and efficiency. Use pure Opus for tasks under 500 LOC.

Best Practices

Use `hooks.PostToolUse` to auto-run prettier
Before: Manual formatting was tedious and error-prone. After: Automated formatting saves time and ensures consistency. Leverage `git-semantic` for team-wide search
Without this: Redundant API calls slow down queries. With this: Faster and more accurate code searches, reducing context-stuffing. Lock in $2/month pricing with ANTHROPIC_BASE_URL
Before: Unpredictable per-token charges. After: Predictable, budget-friendly monthly costs.

Tools & MCP

Vulnetix VDB — Provides real-time security scanning for package dependencies — catch vulnerabilities as you code Tandem — Adds real-time document review to Claude Code — eliminates traditional prompting back-and-forth Forge — Turns Claude Code into a multi-model swarm — orchestrates AI agents for complex tasks

Multi-Agent Patterns

Plan with Opus, Execute with Codex
Use pure Opus for planning tasks under 500 LOC, switch to Codex above 800 LOC for cost efficiency. Tool Search
Defers MCP tool definitions, using a single search tool on-demand — preserves context window for work. Persistent AI Assistants with a2a
Use a2a CLI for persistent memory and skill management in Claude Code projects, enhancing production readiness.

Community Requests

• Native MCP server benchmarking tool
• Improved CLAUDE.md rule enforcement