The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is fundamentally designed to ingest and process user-provided prompts for analysis and optimization, creating a surface for indirect prompt injection.
Ingestion Points: User-defined prompt files (e.g., my-prompt.txt, prompt-v1.txt) and interactive template inputs.
Boundary Markers: The skill explicitly recommends using XML tags (e.g., <instructions>, <context>) to delineate different parts of a prompt, which helps mitigate but does not fully eliminate injection risks.
Capability Inventory: The skill involves executing local Node.js scripts and communicating with the Anthropic API via an external SDK.
Sanitization: The instructions focus on structural separation using XML tags rather than automated sanitization of user input.
[COMMAND_EXECUTION]: The skill provides explicit instructions for the user to execute local Node.js scripts included in the skill's package.
Evidence: node scripts/score-prompt.js my-prompt.txt, node scripts/test-prompt.js prompt-v1.txt prompt-v2.txt test-input.txt, and node scripts/generate-template.js.
[EXTERNAL_DOWNLOADS]: The skill directs users to install an external dependency from a public registry.
Evidence: npm install -g @anthropic-ai/sdk. This targets a well-known service and is documented for transparency.
[CREDENTIALS_UNSAFE]: The skill instructs the user on how to set their API key as an environment variable.
Evidence: export ANTHROPIC_API_KEY="your-api-key". While this is standard practice for developer tools and uses a placeholder, it reminds users that the skill requires access to sensitive credentials to function.

engineering-prompts