The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes an orchestration architecture where a primary agent (orchestrator) delegates tasks to a child implementer agent via a 'brief'. This architecture creates a surface for indirect prompt injection if the brief contains instructions from untrusted data sources.
Ingestion points: The implementer receives instructions through a brief.txt file on stdin, which the orchestrator populates from user input and chat context (referenced in SKILL.md and references/writing-the-brief.md).
Boundary markers: The documentation suggests using XML-style tags like <task>, <repo_constraints>, and <action_safety> to delimit context within the brief.
Capability inventory: The implementer agent (launched via the claude CLI) is provided with tools for file system manipulation (Read, Edit, Write) and shell command execution, according to references/dispatch-and-poll.md.
Sanitization: The strategy relies on human-in-the-loop or orchestrator review of the implementer's generated diffs and test results before committing code changes.
[COMMAND_EXECUTION]: The skill provides instructions for the orchestrator to execute a Node.js script using the command: node "<skill-dir>/scripts/relay.mjs". This involves executing a script from the local file system using a path derived at runtime.
[EXTERNAL_DOWNLOADS]: The skill references a community-maintained GitHub repository (amElnagdy/delegate-skills) as the source for the relay.mjs script and orchestration logic. This repository is not listed as a trusted vendor.
[NO_CODE]: The skill consists entirely of Markdown documentation and reference guides. It does not ship with any bundled executable scripts or automation code, which mitigates immediate execution risk but creates dependency on external components.

claude-delegate