Agent Skills Reached 3.8 Million Files—and Half Were Exact Copies

Giuseppe Destefanis, Daniel Graziotin, Matteo Vaccargiu and Marco Ortu introduced GitSkills on August 11, 2026. The GitSkills preprint describes 3.8 million SKILL.md file occurrences—3,797,117 in total—collected from 282,200 public GitHub repositories in July, grouped into 1,877,981 distinct contents; 50.5% were classified as verbatim copies.

The release offers an unusually large view of the natural-language instructions used to extend AI agents, but not a census of the entire ecosystem. Its headline total reflects GitHub code-search coverage, principally searchable files on public repositories’ default branches, and includes some filename matches that may not satisfy the formal Agent Skills format.

One corpus has two measures of scale

The 3.8 million headline counts file occurrences, not independently authored instructions. GitSkills preserves each discovered file’s repository, path and content hash, then groups files whose contents are byte-for-byte identical.

The gap between the cited occurrence and distinct-content totals is 1,919,136 redundant occurrences. That arithmetic explains why slightly more than half of the files are copies: each redundant occurrence adds another repository path for content already represented by the same hash.

“Exact” has a narrow technical meaning here. A matching hash establishes that two SKILL.md files contain the same bytes; it does not identify the original author, reveal how the text moved between repositories or prove that the surrounding skill directories are interchangeable.

Two repositories can carry identical instruction files alongside different scripts, references or assets. Conversely, files separated by a one-character edit belong to different content groups even when their instructions are functionally similar. The duplication rate therefore measures verbatim reuse, not semantic similarity.

The release separates occurrences from enriched representatives

The public GitSkills dataset card lists 3,797,117 artifact records, 282,200 repository records, 7,264,865 sibling-file records, history coverage for 458,548 files and seven recorded mining runs. The Parquet mirror complements the self-contained SQLite release described in the paper.

The artifacts table is the central index. Each row represents a discovered filename match and records its repository, path, basename, location class and content hash. One representative from each distinct-content group receives additional fields such as the complete text, parsed front matter and body size.

Repository records supply context including ownership, primary language, fork status, licensing and activity dates. Sibling-file records describe material stored beside an enriched representative, allowing researchers to observe whether that particular copy has scripts or reference material.

That representative model creates an important boundary. Folder contents and commit history belong to the selected repository copy, not automatically to every occurrence sharing its SKILL.md hash. An ecosystem study can count textual replication from the hash groups, but it must inspect other copies before making claims about their bundled code or development history.

Why copied instructions can affect agent behavior

The Open Agent Skills specification defines a skill as a directory containing a required SKILL.md file with YAML front matter and Markdown instructions; optional directories may hold executable scripts, references and assets. The required metadata includes a name and a description of what the skill does and when it should be used.

In compatible implementations, the description helps determine whether a skill is relevant to a task, while the instruction body supplies the procedure after activation. Repeated wording can therefore propagate both selection cues and operational directions across repositories, making copied skill text potentially more consequential than duplicated documentation.

GitSkills does not establish that every indexed file was activated, executed successfully or accepted by a particular agent. Nor does an exact textual match show whether the copy came from manual reuse, a template catalog or an automated scaffolding tool. Those are questions for subsequent behavioral and repository-history research.

Public default-branch coverage limits the headline total

The collection pipeline used GitHub code search and related APIs, dividing searches into smaller file-size ranges to work around result limits. That method recovered far more files than a single filename query exposed, but it retained the boundaries of GitHub’s search index.

The main constraints are:

• Private repositories and internal enterprise code are outside the dataset.
• Files found only on non-default branches are not covered by the search method.
• Oversized files and repositories outside GitHub’s indexing eligibility can be missed.
• Fork indexing depends on GitHub’s treatment of the fork relative to its parent.
• Filename matching can include lowercase or compound names that predate, or do not conform to, the specification.

These limits work in both directions. Missing repositories, branches and unindexed files make the corpus a lower bound for potentially relevant public and private artifacts, while broad filename matching introduces records that may be false positives for formal agent skills.

Claims based on GitSkills therefore need an explicit denominator: all discovered occurrences, distinct byte sequences, exact SKILL.md basenames, standard directory locations or files with valid front matter. The dataset establishes that millions of skill-like files are visible in searchable public GitHub code and that verbatim repetition accounts for just over half of them. It does not yet determine how many distinct skills agents actively use, which templates generated the largest copy clusters or how much adoption remains outside public default branches.

Also read:

• How to Use ChatGPT Projects for a Startup Without Mixing Files, Instructions or Memory
• GitHub’s AI Agent Tsunami: 275 Million Commits a Week, 14 Billion Projected for 2026 — And the Platform Is Starting to Crack
• Agentic Marketing Revolution: Okara’s AI CMO Agent Hits 10 Million Views and Takes Down Its Own Infrastructure