Analysis of Copilot Instructions

A reproducible analysis was conducted on 28,721 public repositories collected via the GitHub API, of which 7,755 (27.0%) had Copilot instructions. The full corpus consisted of 165,063 files and 3.3 million directives.

The key finding is that a significant portion of instruction files aren't actual instructions.

• Of a median 50 items per file, only 12 were actual directives.
• Only 27% of total content was directive, while the remaining 73% was headings, explanations, and examples.
• About two-thirds of all directives were abstract expressions that didn't specify a particular tool or file name.

Configuration approaches also varied significantly by agent.

• 37% of projects configured 2 or more agents together.
• The most common combination was Claude + Codex.
• Cursor had a higher proportion of rule files, while Codex and Gemini were centered on base config.
• Subagents were the least substantive, with a median of 61 items, 17 directives, and 17% specificity.

Community skills and role prompts showed similar patterns.

• frontend-design spread to 271 repositories, but specificity was only 2.8%.
• next-best-practices was found in only 76 repositories, but specificity was 92.6%.
• The conclusion is clear: specifically named instructions were more useful than more verbiage.

The dataset and classification results were released as repos.jsonl, stats_public.json, and validation_key.csv, and can be verified directly with npx @reporails/cli check.