The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains multiple instructions that explicitly direct the AI agent to bypass human-in-the-loop approvals and act with full autonomy.
Evidence: "The stage boundaries are checkpoints you satisfy, not approvals you wait for."
Evidence: "The Stage 1 and Stage 2 blocking approvals do not apply on this path: you do not wait for sign-off before implementing. Build autonomously."
Evidence: "NEVER fall back to a blanket approval gate."
[PROMPT_INJECTION]: The instructions guide the agent on how to bypass interactive safety confirmations in the CLI tool being used.
Evidence: "Destructive commands use --force (never --yes; see docs/design/safety.md §3.2)."
Evidence: "--force exists precisely to skip gcx's interactive confirmation, so on this path your question is the only confirmation left."
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and interpret data from external, potentially untrusted sources which could contain malicious instructions.
Ingestion points: Reads OpenAPI specifications, Terraform provider schemas, external Go client libraries, and vendor API documentation (SKILL.md, Stage 1).
Boundary markers: No explicit instructions were found to treat external documentation or code comments as untrusted data or to ignore embedded instructions within them.
Capability inventory: The skill uses bin/gcx for network probes and mise run gate for executing build tasks and tests (SKILL.md, Stage 3 and 4).
Sanitization: There are no documented steps for sanitizing content retrieved from OpenAPI specs or source code before it is used to generate new Go code.
[COMMAND_EXECUTION]: The workflow involves the execution of local binaries and task runners that interact with the file system and network.
Evidence: Execution of bin/gcx for API probing and smoke testing.
Evidence: Execution of mise run gate for linting, testing, and building code.

add-provider