The Second Time Is a Bug Blog

Research data

Data: What 1,866 Popular GitHub Projects Tell Their AI Coding Agents

The scripts, results and AI-reviewed samples behind the study of agent instruction files in 1,866 popular GitHub projects, collected on 15 October 2026.

Download the data (zip, 1 MB)

What is in the download

  • Scripts that list each repository's files at its latest commit, save its agent instruction files, and run the checks and aggregates in the post.
  • The candidate list: 2,424 projects from the GitHub Ranking lists, of which the study keeps the 1,867 with a commit in 2026.
  • Per-project results: the commit read, the instruction files found, the main file and its length, and the analyser's raw flags.
  • Every line the analyser read as a rule, unreviewed.
  • Two reviews by a separate AI reviewer, independent of the analyser: 200 and 300 randomly sampled lines classified one at a time, and a check of every flagged dead path and script against the repository. No person checked every label.

Read this before reusing it

  • The analyser's raw flags overstate problems. Most automatically flagged dead paths and almost all dead scripts were false alarms, so the post reports only the figures the reviewer confirmed.
  • About one line in seven that the analyser reads as a rule is a fragment or a description. Rule-level figures in the post come from the reviewed samples only.
  • The projects' own instruction files are not included. Each can be fetched from its repository at the commit recorded in the results.

The README in the download describes every file. Questions or corrections are welcome through the companion repository.