The Agent Skills Directory

[METADATA_POISONING]: The skill provides highly misleading information in its name and description, claiming compatibility with non-existent AI models such as Claude 5, Opus 5, Sonnet 5, and Fable 5. It also claims a 1M token context window is standard for these models, which does not match the actual capabilities of the current Anthropic model lineup. This deception can lead a user or an agent to misjudge the available resources and context limits of a session.
[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface where untrusted data could influence agent actions. It ingests untrusted user input via $ARGUMENTS for the set, model, and auto commands, which are then used to update local configuration files. There are no boundary markers or delimiters defined in the instructions to separate input data from internal logic. The skill possesses capabilities to use Bash, Read, and Write tools to interact with the file system (specifically ~/.claude-context-optimizer/config.json). No sanitization or validation of the input tokens or model names is specified before they are interpolated into the JSON configuration files.

cco-budget