The Agent Skills Directory

[DATA_EXFILTRATION]: The skill implements strong privacy guardrails. It explicitly forbids copying private corpuses or sensitive source bodies into configuration or chat logs. It mandates the use of sensitivity levels (public, internal, confidential, restricted) to manage data exposure.
[CREDENTIALS_UNSAFE]: The skill instructions explicitly state 'Never store credentials' and include guardrails to 'Stop on a... secret-like value.' It manages references to data sources rather than the data itself.
[COMMAND_EXECUTION]: The skill interacts with a local CLI tool (pragman). The workflow includes mandatory security checks such as pragman --version --json compatibility verification, digest-bound previews, and explicit user approval before any write operations are executed.
[EXTERNAL_DOWNLOADS]: The skill mentions a Node.js package (@prag-man/pragman-exp) for users who lack the required CLI. This package belongs to the skill's own namespace and is offered for manual installation only; the skill is forbidden from automatic installation.
[PROMPT_INJECTION]: No prompt injection or bypass patterns were identified. The instructions are focused on structured context management and maintaining security boundaries between different organizations or projects.

pragman-workspace