The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes past agent conversations to identify user preferences and working styles, creating a surface for instructions embedded in previous history to influence the agent.
Ingestion points: Consumes data from the agent-transcripts/ directory within the active workspace.
Boundary markers: The instructions do not define explicit delimiters or boundary markers when reading raw transcript content during the mining process.
Capability inventory: Orchestrates the skill-creator tool, which has the capability to write new or modified skill files to the .agents/skills/ directory.
Sanitization: No explicit sanitization or filtering of the ingested transcript content is specified before using the derived patterns to draft new skills.
Context: The risk is mitigated by explicit guardrails forbidding the use of globbing across different workspaces and the core focus on identifying recurring patterns across multiple 'slices' of history before codifying them.

automate-me