Split files generated by --skill-generate into smaller disk files to reduce loaded context size.

[whinc]

Description

I like the new --skill-generate feature after trying it out, but I noticed an area that could be improved. The generated files.md in the skill directory currently contains the full contents of all files, which can cause the model context to become excessively large. If only the necessary files were loaded progressively into the model context, the skill would be more focused and effective.

This is a real example:

ls -lh .claude/skills/formily/references          
total 3008
-rw-r--r--  1 whincwu  staff   1.5M Dec 15 11:23 files.md      <-- too large
-rw-r--r--  1 whincwu  staff   8.6K Dec 15 11:23 project-structure.md
-rw-r--r--  1 whincwu  staff   2.6K Dec 15 11:23 summary.md

My suggest is split files.md into more smaller disk files or chunks to reduce the loaded model context size. Maybe support a new cli argument to enable this behavior.

[coderabbitai]

📝 CodeRabbit Plan Mode

Generate an implementation plan and prompts that you can use with your favorite coding agent.

• Create Plan

Examples

• Example 1
• Example 2

🔗 Similar Issues

Possible Duplicates

• Feature Request: Automatic Chunking (a.k.a. “YOLO Divide”) for Oversized Codebases #424

Related Issues

• Add --summary CLI option for LLM-based code summarization #511
• Support generate claude skills package #953
• Support generating claude skills package #952

👤 Suggested Assignees

• whinc
• rockmandash
• yamadashy

---

🧪 Issue enrichment is currently in early access.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

[yamadashy]

Hi, @whinc !
Thank you for proposing the --skill-generate feature and continuing to provide valuable feedback!

I actually considered splitting files during the initial implementation, but I decided to keep everything in a single files.md for the following reasons:

• The files.md is not intended to be loaded entirely into the context. Instead, it's designed for agents to reference only the necessary parts using Grep or Read tools with line number specification.
• The SKILL.md file includes instructions on how to reference files this way, and it's been working well with Claude Code.
• Splitting into multiple files would require additional indexing or metadata to help agents determine which file to read, potentially adding complexity.

That said, I'm open to hearing more about your use case. If you're experiencing issues with a specific tool or environment, please let me know and I'd be happy to explore potential improvements.

[whinc]

I agree with you. After using the skills generated by repomix --skill-generate for a while, they behave exactly as you described: instead of loading the entire files.md, the model searches for only the necessary snippets, with at most an initial peek of the first 2,000 lines.

[yamadashy]

Thanks for testing it out and sharing your feedback!
I'm glad the skill is working as expected.