The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted qualitative data (interview transcripts) from local file paths provided by the user. Maliciously crafted transcripts containing hidden instructions could potentially influence the agent's behavior during the multi-stage coding process.
Ingestion points: The skill reads qualitative data (PDF, Word, TXT) and previous coding tables from user-specified paths.
Boundary markers: While the skill uses structured Markdown tables and sequence tagging (e.g., S1, S2) for analysis, it lacks explicit instructions to the model to ignore or escape embedded instructional content within the data.
Capability inventory: The agent has the ability to write Markdown and Excel files to the local directory and execute shell commands to run Python scripts.
Sanitization: Raw interview text is processed and interpolated into analysis prompts without explicit sanitization for command or instructional patterns.
[COMMAND_EXECUTION]: The skill uses a helper script scripts/export_coding.py to aggregate results into an Excel file. This script is executed via a shell command at the end of the workflow. The script also includes a routine to install the openpyxl library using pip if it is not already available in the environment.
[EXTERNAL_DOWNLOADS]: The skill performs a runtime dependency check and installs the well-known openpyxl package from the official Python package registry (PyPI) to support Excel generation.

grounded-coding