I've gone through a few iterations of my Claude Code setup. The first version had too much going on: multi-agents running in the background, Ruflo for automated tasks, Graphify for knowledge graph generation. Sessions got expensive. I hit usage limits fast and too much of that compute went to things that weren't helping me ship.
So I did a clean reinstall. Here's what I kept and why.
What I Removed
Multi-agents. I had a custom multi-agent setup for orchestrating parallel tasks. It sounded great on paper but the token overhead was significant. Most tasks don't need three agents working simultaneously — one focused agent does the job with less waste.
Ruflo. Ruflo automates repetitive Claude workflows. Useful in theory, but it added complexity and token burn for sessions where I just needed Claude to write or review code.
Graphify. Graphify builds a knowledge graph of your codebase — it maps function calls, dependencies, and communities of related code. The output is genuinely useful for understanding large repos. But it's expensive to generate and maintain. I removed it from my active setup and now run it manually when I actually need the graph.
What I Kept (and Why)
Caveman Mode
This is a communication style skill. When active, Claude drops filler words, articles, and unnecessary hedging. Responses get compressed significantly — sometimes 70–75% shorter without losing technical content. For a developer reading output in a terminal all day, this matters. I run it at ultra level, which also uses arrows for causality and abbreviates common terms.
Claude-Mem
This is the biggest one. Claude-Mem is a persistent memory plugin by thedotmack (currently at v12.1.1). It gives Claude a memory layer that survives across sessions.
It works through hooks that fire at key lifecycle points:
- SessionStart — loads context from previous sessions
- UserPromptSubmit — initializes session tracking
- PreToolUse — loads file-specific context when reading code
- PostToolUse — records observations after tool use
- Stop / SessionEnd — summarizes the session and writes it to memory
Under the hood it uses a local Chroma vector database for semantic search across past observations. The model it uses for memory operations is claude-sonnet-4-6. It also exposes an MCP server (bun-based) for search queries.
The practical result: Claude knows what we've worked on before. It remembers decisions, context, and patterns without me having to re-explain every session.
Karpathy Guidelines (andrej-karpathy-skills)
This skill encodes behavioral guidelines derived from Andrej Karpathy's observations about common LLM coding mistakes. The core ideas: don't overcomplicate, make surgical changes, surface assumptions instead of silently guessing, and avoid rewriting things that aren't broken.
It doesn't change what Claude can do — it shapes how Claude approaches problems. Less drift, more precision.
ECC Tools Cost Audit
ECC is a specialized skill for auditing Claude Code's own cost and behavior. It helps track things like runaway tool calls, premium model leakage, and quota usage patterns. Useful for staying aware of what sessions actually cost.
RTK (Rust Token Killer)
RTK is a CLI proxy that compresses terminal output before it reaches Claude. When you run git log, cargo test, or pnpm list, RTK intercepts the output and strips everything that doesn't matter: verbose metadata, ASCII art, redundant lines. The result is 60–90% token reduction on command output.
The skills rtk-tdd and rtk-triage help Claude work with RTK effectively — understanding the compressed format and knowing how to debug using it.
UI/UX Pro Max
This skill is a design reference system. It covers 50+ visual styles, 161 color palettes, 57 font pairings, and patterns for 10 technology stacks including React, Next.js, Tailwind, and shadcn/ui. When I'm building frontend work, this skill gives Claude the vocabulary and patterns to produce opinionated, non-template-looking UI instead of defaulting to generic layouts.
MCP Servers
One MCP server runs for Claude by default: the claude-mem search server. It's a bun process that handles semantic search queries against the memory database.
The other MCP I added is Playwright MCP (MS Edge). This gives Claude browser control — it can navigate pages, take screenshots, click elements, and verify UI behavior directly in Edge. Useful for testing portfolio changes without manually opening a browser.
Rules System
I have rules installed for 15 languages and domains — Rust, TypeScript, Python, Go, Swift, Kotlin, PHP, and more. These shape how Claude writes code in each language: naming conventions, error handling patterns, testing requirements, and what to avoid. The rules live in ~/.claude/rules/ and are loaded as context for relevant sessions.
The Result
Cleaner sessions. Lower cost per session. Claude has persistent memory, compressed output, design intelligence, and browser control when needed — without the overhead of systems that were running more for completeness than for actual use.
The lesson from the first setup: more tools doesn't mean more productive. The question is whether a tool is actively helping you ship or just consuming tokens while it waits to be useful.