The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The skill utilizes the ! command syntax in its documentation to describe how expressions are evaluated and values are collected at runtime. This mechanism allows the execution of shell commands at the time the skill context is processed.
Evidence in SKILL.md: "In inherited Context blocks, run each ! command to collect the named value."
[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using the claude CLI. It uses shell command substitution $(cat <handoff-file>) and interpolates variables derived from user-provided $ARGUMENTS (such as descriptive name and model selection) into the command string.
Evidence in SKILL.md: claude --bg --name "<descriptive name>" --model "<model>" -- "$(cat <handoff-file>)".
[INDIRECT_PROMPT_INJECTION]: The skill involves processing the 'current conversation' and repository state to generate a handoff summary for a new agent. This ingestion of untrusted data from the conversation history creates a surface where instructions embedded in the original text could influence the behavior of the background agent.
Ingestion points: SKILL.md identifies conversation history, repository state, and user arguments as data sources.
Capability inventory: Shell execution via claude and codex background tasks.
Boundary markers: None specified to differentiate between the agent's summary and potentially malicious user-provided text in the context.
Sanitization: No explicit instructions are provided to sanitize or escape the content of the generated handoff file before it is read and executed as part of a command.

splitoff