Even after the instruction files for AI agents were split, one file's character count grew 69% in 12 days

Summary

This is a record of splitting two of this site’s own rule documents, which AI agents read, into a SKILL.md and reference files. SKILL.md is the central file of a skill in Anthropic’s Agent Skills mechanism. A reference file is a file that SKILL.md points to by name and that the agent reads when it needs it. The two rule documents that were split are the writing rule document and the style rule document.

In the split work on 2026-08-08, the combined character count of the two SKILL.md files went from 81,480 to 38,608 characters. That is a 52.6% reduction. The combined total including the reference files grew by 4,375 characters.

On 2026-08-08, the split work was followed by deduplication work and finishing work. The deduplication work kept the text of each rule that appeared in several documents in one document, and replaced the text in the other documents with references. The finishing work was the last task of the 2026-08-08 refactoring. When the finishing work was done, the SKILL.md of the writing rules had 11,536 characters. Twelve days later, on 2026-08-20, it had grown to 19,505 characters. That is a 69.1% increase. Over the same 12 days, the line count of the SKILL.md of the writing rules grew by 29.2%. Both SKILL.md files stayed under 500 lines.

In the split work, an editing agent and a verification agent compared every line before and after the split as a multiset of lines. A multiset comparison of lines compares every line of the original file against every line of all the files in the new layout, counting how many times each identical line appears. The verification agent is an AI agent run separately from the editing agent. The verification agent confirmed that zero lines were lost in either rule document.

In the deduplication work, the verification agent pointed out nearly lost rules and rules whose meaning had changed. The records do not say what method the verification agent used to compare in this work. In the job procedure documents, five rules were found that had nearly been lost because the referenced document did not contain the rule text.

In the body, I first show how many characters the split removed and how many were added over the 12 days. Next, I explain how characters were counted and the byte counts left in the records. After that, I cover which files the model that writes the final draft reads after the split, and how the split work confirmed that no rules were lost. Then I turn to the nearly lost rules that the verification agent pointed out in the deduplication work. I also describe a reproduction experiment that gave only the new layout to a model without context. Following that, I look at the line number references that had drifted even before the split, and their conversion to section name references. The eighth section explains the breakdown of the growth over the 12 days and growth that a line limit cannot detect. Finally, I write about what this record cannot show and six items to check when splitting rule documents.

What you can take away

This article is for people who manage instruction files or skills for AI agents split across several files. You will learn the following four things.

• How to confirm, with a multiset comparison of lines, that no rules were lost in a split
• The pattern of places where rules were nearly lost when duplicated rule text was replaced with references, and what to check to find that pattern
• The number of lines whose line number references drifted through edits, and the number of lines whose section name references pointed to renamed sections
• An example where the character count grew even though the line limit was respected

This article is a record based on this site’s own rule documents and development records.