The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains robust instructions to maintain safety boundaries and includes evaluation cases specifically designed to test resistance against malicious prompts. It explicitly directs the agent to refuse out-of-scope or unsafe requests while following documented contracts.
[DATA_EXFILTRATION]: The skill handles sensitive credentials such as NPM_TOKEN and GITHUB_PERSONAL_ACCESS_TOKEN as part of its setup operations. It includes strict constraints to redact these secrets from logs and uses non-exposing validation methods, such as checking token length rather than printing values. Network operations are directed towards well-known services like the npm registry and CircleCI API for legitimate configuration checks.
[COMMAND_EXECUTION]: The skill utilizes standard repository management and verification commands (e.g., harness init, pnpm check). It follows a 'preview-first' philosophy, requiring dry-runs and user confirmation for all impactful or destructive actions, such as CI migrations or file re-scaffolding.
[INDIRECT_PROMPT_INJECTION]: There is a potential surface for indirect prompt injection as the skill ingests project data (package.json scripts) to generate environment configuration blocks. The skill mitigates this by requiring user checkpoints for any non-autogenerated content and treating user-owned files as protected.
[EXTERNAL_DOWNLOADS]: The skill manages the installation of necessary tooling from the official npm registry. These downloads are central to the skill's purpose and are performed using standard package managers.

coding-harness