The Agent Skills Directory

[DATA_EXFILTRATION]: The skill reads the contents of local files and transmits them to the Anthropic API or the 'claude' CLI for processing. While this involves data transfer to a third-party service, it is the primary intended function of the token-reduction tool. The skill implements a security filter in scripts/compress.py using the is_sensitive_path function to prevent the accidental upload of sensitive files such as SSH keys, environment configurations, and credential files.
[COMMAND_EXECUTION]: The script scripts/compress.py utilizes subprocess.run to call the external 'claude' CLI tool. The command is constructed with a static list of arguments, which prevents shell command injection. This is used as a fallback if the primary Python API client is not available.
[PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection as it ingests untrusted data from local files into an LLM prompt.
Ingestion points: File content is read via filepath.read_text() in scripts/compress.py and passed to the LLM.
Boundary markers: The prompt uses a simple text label without robust delimiters or instructions to ignore embedded commands in the source file.
Capability inventory: The skill has the capability to overwrite local files via filepath.write_text() and execute CLI tools.
Sanitization: The skill mitigates risks through a post-processing validation step in scripts/validate.py, which ensures that the compressed output maintains the original file's structure, including headings, URLs, and code blocks, reverting changes if discrepancies are detected.

genshijin-compress