The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests untrusted user data to generate and update executable Harness pipelines.
Ingestion points: User-provided inputs for flag names, treatments, environment IDs, and custom logic for shell scripts (found in SKILL.md Phases 4 and 7).
Boundary markers: Absent. There are no explicit delimiters or instructions to ignore embedded commands within the ingested data.
Capability inventory: The skill utilizes harness_create and harness_update tools (mapping to SKILL.md) to apply generated YAML configurations which can include ShellScript steps capable of performing network operations and command execution (as seen in references/blueprints/l3-segment-sync.yaml and references/blueprints/l4-test-targeting.yaml).
Sanitization: Not present. The skill relies on the underlying model's safety logic and the explicitly instructed human confirmation step in Phase 8 to prevent malicious generation.
[COMMAND_EXECUTION]: The skill provides blueprints that generate pipelines containing shell script execution steps.
Evidence: references/blueprints/l4-test-targeting.yaml includes an npm run test:e2e command, and references/blueprints/l3-segment-sync.yaml includes a bash script for fetching data.
[EXTERNAL_DOWNLOADS]: The skill includes a blueprint that generates a pipeline using curl to fetch data from an external source.
Evidence: references/blueprints/l3-segment-sync.yaml contains a template using curl -s https://api.example.com/premium-users to fetch user keys.

fme-pipeline