# Claude Code Digest — Mar 22–24

> Agent Teams hit mainstream, Bossa MCP adds persistent memory, and 99.4% of your tokens are input reads — here's how to cut them in half.


Agent Teams hit mainstream, Bossa MCP adds persistent memory, and 99.4% of your tokens are input reads — here's how to cut them in half.
99.4% of Claude Code tokens are input — you read 166x more than you write

## Trending Now
**🔥 Agent Teams Going Production**
Since Feb 5 launch, teams are deploying 3-4 teammate setups for parallel code review (security + perf + tests). Each teammate gets its own 1M context window — powerful but token-expensive.
**🔥 Hooks as the New API**
17 hook events now available. Smart developers use PreToolUse hooks for auto-approval workflows and PostToolUse for auto-formatting — turning Claude into a fully autonomous loop.
**📈 Worktree Isolation for Parallel Dev**
The /batch command handles worktree isolation automatically. ccswarm coordinates specialized agents using Git worktrees for concurrent development on separate branches.
**✨ Context7 Replacing Stale Docs**
Context7 by Upstash pulls current, version-specific docs for thousands of libraries at query time. No more hallucinated APIs from outdated training data.

## Best Practices
**Run Claude in a Container (Worker Pattern)**
Use tmux as the control layer: local Claude controls a sandboxed Claude instance. Long-running tasks, zero risk, full autonomy. The one pattern power users swear by.
**Keep CLAUDE.md Under 200 Lines**
It's read on EVERY turn, not just session start. A 500-line CLAUDE.md burns tokens silently. Move reference content to .claude/docs/ and use skills for domain knowledge.
**Use /compact with Custom Instructions**
Most people just type /compact. Power move: /compact Focus on code samples and API contracts — tells Claude exactly what to preserve during summarization.
**Ask Claude to Ask YOU Questions**
Before any complex task, say: "Before planning, ask me 5 questions about requirements." This one prompt eliminates 40% of misunderstandings and wasted iterations.

## Tools & MCP
**Bossa MCP** — Adds persistent memory to Claude Code. Stop pasting context every session — Bossa remembers project state, decisions, and patterns across sessions.
**Nano Banana** — Wraps Google Gemini image generation. Claude Code can now generate and edit images mid-conversation. Design mockups without leaving the terminal.
**Feature Dev Skill (89K installs)** — The #1 plugin. Invokes a 7-phase workflow: requirements → exploration → architecture → implementation → testing → review → docs. One command.
**ClaudeRank Widget** — Open-source widget showing real-time Claude Code usage stats. Track tokens, costs, and model switching patterns across sessions.

## Multi-Agent Patterns
**Master-Clone Architecture**
Main agent decomposes task, spawns 3-4 subagents with isolated worktrees. Each clone works on separate files. Results merge back. 3x faster, same cost.
**Subagent vs Team Decision Rule**
If workers don't need to communicate → use subagents (cheaper). If they need to share findings and challenge each other → use Agent Teams. Teams use significantly more tokens due to separate context windows.
**Gemini CLI as Fallback Agent**
Create a skill that routes web-access tasks to Gemini CLI when Claude's WebFetch is blocked. Gemini fetches Reddit, paywalled docs, and sites Claude can't reach.

## Community Requests

- Native cost tracking dashboard — current /cost is too basic, developers want per-session and per-task breakdowns
- Better Agent Team debugging — no visibility into teammate context or why decisions were made
- Nested teams support — teammates can't spawn sub-teams, limiting complex orchestration
- MCP server hot-reload — changing .mcp.json still requires session restart


---
*Originally published on [gentic.news](https://gentic.news/article/claude-code-community-digest-mar-22-24-2026)*