The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell scripts and development tools as part of its core workflow. Evidence includes the execution of bash scripts/runbooks/select-runbooks.sh --skill he-implement to load project runbooks. It also executes scaffolding commands such as bun create next-app or cargo init, and performs continuous Git operations and test executions.
[PROMPT_INJECTION]: An indirect prompt injection surface exists because the skill follows instructions from external plan files.
Ingestion points: Reads instructions from docs/plans/active/<slug>-plan.md, docs/generated/README.md, and the output of the runbook selection script.
Boundary markers: The skill lacks explicit delimiters or instructions to treat data from plan files as untrusted, meaning the agent may follow instructions embedded within those documents as if they were system instructions.
Capability inventory: The agent can execute shell commands (bash), manage Git repositories, and spawn subagents for parallel tasks.
Sanitization: No evidence of content validation or sanitization is present before the agent processes and acts on the contents of the plan files.

he-implement