The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a workflow that processes data from potentially untrusted sources, including the project's own architecture and external communication channels.
Ingestion points: In SKILL.md, the orchestrator is described as reading from .adlc/drafts/adr/, .adlc/memory/adr/, and project-level architecture. It also explicitly reads consistency findings from a PR/MR/issue comment bus during its correction loop.
Boundary markers: No explicit delimiters or boundary markers are defined to isolate untrusted data (like code-extracted ADRs or PR comments) from the agent's instructional context.
Capability inventory: The skill has the capability to execute a series of related architectural tools (architect-specify, architect-clarify, architect-implement, architect-analyze) and performs file system writes to AD.md and the .adlc/ directory as specified in SKILL.md and workflow.yml.
Sanitization: The workflow logic in workflow.yml performs direct string matching on command output (steps.analyze.output.stdout | contains(...)) without documented sanitization or validation of the processed content.

factory-architect