The Agent Skills Directory

[PROMPT_INJECTION]: Presence of highly deceptive and misleading metadata within reference documents.
Multiple reference files, including references/architecture-patterns.md, references/code-organization.md, references/deployment-pipelines.md, references/error-handling.md, references/performance-optimization.md, references/security-best-practices.md, references/state-management.md, and references/testing-strategies.md, contain hundreds of identical sections of filler text.
This content discusses kernel-level optimizations (e.g., epoll_wait, io_uring, eBPF) that are irrelevant to data versioning, appearing designed to artificially inflate the skill's complexity or evade simple static analysis by overwhelming the context.
[COMMAND_EXECUTION]: The skill facilitates the assembly and execution of shell commands using external data inputs without sanitization.
File: SKILL.md and references/lakefs-patterns.md.
The skill instructs the agent to use user-provided strings for branch names and commit messages in commands such as lakectl branch create, lakectl commit, and dvc commit -m.
The lack of explicit sanitization or validation instructions for these inputs creates an indirect prompt injection surface where a malicious user could potentially execute arbitrary commands via shell metacharacters.
[EXTERNAL_DOWNLOADS]: Reference to external tool installation and remote resource fetching.
File: references/dvc-patterns.md and references/lakefs-patterns.md.
Instructs on installing dvc and dvc-s3 via package managers and utilizes Docker images for services like LakeFS and Nessie from official registries.
[SAFE]: Proper credential handling in configuration examples.
File: references/dvc-patterns.md.
The documentation correctly demonstrates using environment variable placeholders (e.g., ${AWS_ACCESS_KEY_ID}) for sensitive secrets rather than hardcoding them into the configuration files.

data-data-versioning