Is my Claude Code workflow overengineered? Looking for ways to reduce token usage
retained text
I use Claude Code to maintain a WordPress site built with Etch and Automatic.css. I define requirements and make decisions; CC implements them. Codex acts as a separate reviewer and files findings as tickets. My setup uses two private Git repos: **Project repo:** code, Markdown tickets, specifications, ADRs, project-specific skills, and validation scripts. **Shared Wiki repo:** reusable skills, technical references, and known mistakes organized by language/tool. Skills have short core instructions and optional references. Each task goes through custom /ticket-start and /ticket-done workflows: read relevant context, implement, verify acceptance criteria, update the ticket, record reusable lessons, commit, and push. Larger changes get independent review. I also use [Matt Pocock’s skills](https://github.com/mattpocock/skills), mainly TDD and code review. This helps prevent repeated mistakes, but usage feels roughly **2× higher**. That isn’t a controlled comparison yet. Initial context only increased from about **81k to 84k tokens**, while finishing tasks accounted for **21–51% of estimated cost** in a small sample. I suspect the overhead comes from repeated reading, reviews, and updating/validating knowledge across both repos. For people using similar setups: What would you simplify first? How do you keep shared knowledge useful without loading too much context? When does a mistake deserve a permanent rule? Do you review every task independently, or only risky changes? Would you keep custom workflows or use a thinner layer around Matt’s skills? I want to keep reliable verification and reusable knowledge while reducing the cost of the process itself.
document fetched 2026-10-30 00:43 · canonical https://reddit.com/r/ClaudeCode/comments/1xih1v4/is_my_claude_code_workflow_overengineered_looking