The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process user-provided 'draft prompts', which represents an indirect injection surface. However, the instructions incorporate robust boundary markers, explicitly commanding the agent to remain in an advisory-only mode and providing clear protocols to reject implementation or execution requests. These guardrails effectively mitigate the risk of the agent being manipulated by malicious instructions embedded within the analyzed data.
Ingestion points: The skill ingests untrusted text via the user's draft prompt during the analysis pipeline.
Boundary markers: The skill contains multiple explicit directives such as 'Advisory only — do not execute the user's task' and 'do not switch into implementation mode inside this skill'.
Capability inventory: The skill performs file-read operations (Phase 0) on project configuration files (e.g., package.json, go.mod) to detect the environment, but does not possess network-write or command-execution capabilities within its own scope.
Sanitization: The skill relies on behavioral constraints and user-redirection instructions to handle unsafe interpolation or task-switch attempts.
[SAFE]: The project detection logic (Phase 0) accesses standard, non-sensitive project metadata files (like package.json, go.mod, and Cargo.toml) to provide contextually relevant recommendations. This is a standard and expected behavior for developer-oriented tools and does not involve accessing sensitive user credentials or system files.

prompt-optimizer