Enhancing Claude Code Efficiency & Reliability

TITLE: Claude Code tuning: context collapse + token burn I run Claude Code on 3 private codebases, RN, react and backend every day, as my main way of building. It works, but it is expensive and unreliable in specific, measurable ways. I want someone to fix that. These numbers come from my own session transcripts: - 25,000-40,000 tokens of process documents are read before any actual work starts. One of those files is 117 KB. There are 27 of them in total. - 52 forced auto-compactions across 6 sessions. Individual sessions have run 7-14 days without ever ending. - 1,283 images were loaded into a single conversation as raw pixel data. - 42% of the time a file is opened, it had already been opened earlier in that same session. THE FIRST PROBLEM My process documentation is strangling my own agent. Rules, handoff notes and work ledgers have grown into a pile that gets read from the top every time a session starts, before anything useful happens. I need it restructured so the agent reads what applies to the task in front of it and nothing else. You will not see the contents. You will see sizes, structure and load order, and tell me how to reorganise it. AFTER THAT - Sessions that use intelligent compaction - Agents that finish a task instead of stopping halfway - Output written in plain English by default - Work that reflects the big picture instead of missing context and getting rejected I know the words hooks, subagents, MCP, skills, output styles and settings exist. I do not know which of them actually solve the above. I am not prescribing anything. If your answer is "do not do X, do Y", say why. I have already been told that installing a code index or a graph tool is the answer, and my own measurements say it would save me single digits. Convince me otherwise with numbers or propose something better. HOW WE WORK The codebase is private and stays private. No repo access, no NDA to sign, nothing shared. We work entirely over screenshare: you direct, I drive and type. That means you are teaching while you configure, which is what I want. At the end I understand and own my own setup. If that is a dealbreaker, no hard feelings, but please do not apply. RATE AND HOURS Hourly, ongoing. Roughly 2-3 hours for the initial setup, then 2-3 hours a week after that. Think expert on retainer, not a project. Budget $15-20/hr depending on experience. TO APPLY Answer these five. If you do not answer them specifically, I will archive without reading. 0 - MOST IMPORTANT propose around 10 skills, plugins and hooks (harnesses), etc that WORK FOR YOU in order to make the whole setup perform much better 1. How would you restructure a large pile of project rules and handoff documents so an agent loads only what the current task needs? Be concrete about the mechanism. 2. What would you change in the first two hours, in order? For each: what it changes, and what number I would watch to know it worked. 3. Long-running sessions and auto-compaction: should I be preventing compaction, or ending sessions? Why? 4. My agents stop mid-task. What actually causes that? Which causes are fixable by configuration, and which by how I write prompts? 5. What would you deliberately NOT install, and why?