My Claude Code Agent for Writing Prompts (olshansky.info)

The author introduces a "Prompt Writer Agent" built for Claude Code that automates the tedious but essential work of maintaining and tuning prompts. Instead of treating prompts as one-and-done, the agent captures the end of productive sessions (using Claude Code’s UI), turns past conversations into reusable slash commands via Claude’s /resume feature, writes or updates prompt files under ~/.claude/agents/prompt-writer.md, and integrates with a simple git workflow (cd ~/.claude; rungit diff .; inspect edits; commit and push). The author uses it to convert session fragments—like cross-architecture release-build wrangling—into versions of prompts that are immediately reusable. This matters because prompts are living documentation: keeping them current improves reproducibility, reduces cognitive overhead, and enables teams to iterate on LLM-driven workflows more reliably. Technically, the agent leverages Claude Code’s UI strengths, works alongside a broader personal LLM stack (ChatGPT, Codex, Gemini, Copilot CLI, llm CLI, LM Studio), and fits into familiar dev tooling (editors, git). The repo is public and the author plans to add more curated prompts, encouraging users to star the project and adopt an automated, versioned approach to prompt engineering.