The Agent Skills Directory

[PROMPT_INJECTION]: The instruction block at the beginning of the file uses deceptive language ("This prompt was injected into your context") and imperative overrides ("STOP -- DO NOT READ THIS FILE") to attempt to bypass the agent's standard operating procedure of reading its instructions.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to execute a command-line utility named mempalace. It performs filesystem operations including directory creation (mkdir), file writing (cat), and copying project artifacts (cp) into a staging area (.planning/.mempalace-stage).
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external project files which may contain untrusted data.
Ingestion points: The skill reads CONTEXT.md, PLAN.md, and SUMMARY.md based on the execution phase or $ARGUMENTS in SKILL.md.
Boundary markers: There are no boundary markers or delimiters used to separate the content of these files from the instructions passed to the tool.
Capability inventory: The skill has the ability to write to the filesystem and execute shell commands through the Bash tool.
Sanitization: The content of the artifacts is copied verbatim into the staging directory without any validation or sanitization, potentially allowing malicious content within those files to influence the mempalace tool's behavior during the mine operation.

gsd-mempalace-capture