The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the workspace, including project briefs and memory folders, to drive the failure analysis.
Ingestion points: Step 1 in SKILL.md performs a workspace scan for CLAUDE.md, memory/ folders, and project briefs.
Boundary markers: The skill uses a scratch file (premortem-context.md) to pass context to sub-agents, which acts as a structured boundary, although no explicit "ignore instructions" delimiters are used for the ingested content.
Capability inventory: The skill can write files to the workspace (premortem-context.md, premortem-report-[timestamp].html) and spawn multiple parallel sub-agents.
Sanitization: No explicit sanitization or validation of workspace content is documented before it is processed by the model.
[COMMAND_EXECUTION]: The skill utilizes parallel sub-agent orchestration (model: "sonnet") to perform deep-dive investigations into failure reasons. This is a core feature of the agentic workflow and is used here for narrative synthesis and analysis.
[DATA_EXFILTRATION]: While no network operations are present, the skill identifies and reads potentially sensitive project metadata (CLAUDE.md, project briefs) to build the premortem context. This is consistent with the stated purpose of analyzing a project plan.

premortem