The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from Metaflow run logs, error messages, and stack traces, which serves as a vector for indirect prompt injection.
Ingestion points: Metaflow run status, failing step details, and error summaries are retrieved in step 3 of the workflow and formatted into the final report in step 5.
Boundary markers: The instructions do not define specific delimiters or "ignore instructions" warnings to isolate the untrusted log content from the agent's core logic.
Capability inventory: The skill possesses the capability to invoke the debug-run skill and perform network operations to upload HTML reports to Amazon S3.
Sanitization: The skill mitigates risk by instructing the agent to use deterministic math for classification and limiting the LLM to prose narration only, which discourages the agent from following instructions embedded in the data.
[DATA_EXFILTRATION]: The skill is designed to transmit report data to an external cloud storage service.
Evidence: Workflow step 6 and the self-review checklist specify that the HTML version of the report is uploaded to Amazon S3 using a shared delivery pattern.
[COMMAND_EXECUTION]: The skill dynamically invokes another agent skill based on data retrieved during the workflow.
Evidence: Step 3 of the workflow instructs the agent to invoke Skill(skill="debug-run", args="<FlowName>/<run_id>") using flow names and run IDs obtained from the Metaflow environment.

ml-pipeline-report