The Agent Skills Directory

[Indirect Prompt Injection Surface]: The skill is designed to ingest and process external data sources, such as session transcripts and corpora, to derive persona behaviors. While this introduces a potential surface for indirect prompt injection if the source data contains adversarial instructions, the skill includes multi-tier synthesis and corroboration steps intended to filter noise and quarantine non-authentic vocabulary.
Ingestion points: Processes external transcripts and session data in Step 4.
Boundary markers: Mentions excluding automated noise and eval-harness artifacts in Step 4b.
Capability inventory: Utilizes bash (jq/grep/sed), python3, and amplifier CLI tools for data extraction and skill validation.
Sanitization: Employs two-tier synthesis (Step 4d) to rank traits by corroboration and exclude AI-authored vocabulary.
[Local Command Execution]: The skill workflow involves executing local shell commands and utility scripts to manage data and verify the generated output.
Evidence: Uses bash for data extraction (Step 4b), python3 for metadata length verification (Step 7), and a local CLI tool (amplifier) for running and updating skills (Steps 6, 10). These commands are integrated into the development lifecycle and appear consistent with the skill's authoring purpose.

personafy