I need Claude.ai to generate commit messages, PR descriptions, and even some documentation that matches our internal style guide. The guide is about 15 pages of detailed rules, examples, and brand voice requirements.
The obvious approach is to paste it into the context window, but that burns a huge chunk of my tokens on every single prompt. I've tried referencing it with "according to our style guide..." but the results are inconsistent. Claude doesn't know the specifics unless I paste them in each time.
What's the working solution here? I'm looking for something reproducible, not a one-off hack.
My current ideas:
* **Create a distilled "Claude-ready" version:** Boil it down to a one-page checklist of absolute rules. Less nuance, but more likely to fit.
* **Use a vector store/RAG setup:** This seems overkill for a static document, but maybe necessary. Feed the guide into a system that retrieves relevant snippets.
* **Leverage Claude's file upload:** Upload the guide as a `.txt` file and try to point to it conversationally. I'm skeptical about how well it references uploaded files across a long chat.
Has anyone built a reliable pipeline for this? I care about consistency across multiple team members and sessions. If the solution is "use the API with a system prompt," I need to know the practical limits of that system prompt before it starts ignoring instructions.
Build once, deploy everywhere