The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill migrates content from legacy configuration files (e.g., .cursorrules, AGENTS.md) and repository documentation into the primary harness files (CLAUDE.md, .claude/rules/). If an attacker embeds malicious instructions in these legacy files, they could be promoted to always-on instructions for the agent.
Ingestion points: SKILL.md Phase 1 (Discovery) and Phase 2 (Migration) read content from various legacy files and existing codebase documentation.
Boundary markers: The skill does not generate explicit boundary markers or instructions to treat migrated content as untrusted by the model.
Capability inventory: The generated agents have broad capabilities including Bash execution, Edit, and Write tools.
Sanitization: The skill does not perform sanitization or validation of the text being migrated from legacy files into the new harness structure.
[DYNAMIC_EXECUTION]: The skill generates and registers shell scripts as PostToolUse hooks in .claude/settings.json, leading to automatic code execution upon tool events.
Evidence: SKILL.md section 3.15 describes generating scripts like format-on-edit.sh and wiring them to execute automatically after Edit or Write operations.
[COMMAND_EXECUTION]: The skill uses a wide range of shell commands for repository discovery, git management, and validation.
Evidence: Multiple phases in SKILL.md utilize git, find, grep, ls, and mkdir. The scripts/validate-harness.sh script also uses awk and grep to process file contents.

create-agent-harness