The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Multiple workflow components fetch external data. The mcp-pipeline-builder clones repositories from arbitrary URLs, and the article-evaluation-pipeline and research-to-article skills fetch content from the web using WebFetch or curl.
[REMOTE_CODE_EXECUTION]: The skill facilitates the creation and execution of dynamic content. Specifically, the hook-development-pipeline executes newly written Python scripts to verify performance metrics, and the pipeline-test-runner executes generated skills against test targets. The mcp-pipeline-builder generates entire project structures and runs build commands like npm install and npm run build on generated assets.
[COMMAND_EXECUTION]: The workflow references frequently execute shell commands for development tasks, including running linters (golangci-lint, ruff, tsc), build tools, and system-level configuration validations (nginx -t, systemctl reload). It also modifies configuration files (settings.json, ~/.claude.json) to register new hooks and MCP servers, establishing persistence for generated components.
[PROMPT_INJECTION]: The skill has a large attack surface for indirect prompt injection as it processes untrusted data from multiple sources to influence its complex generation and execution logic.
Ingestion points: references/mcp-pipeline-builder.md (source code from external repos), references/article-evaluation-pipeline.md (arbitrary web content), references/github-profile-rules.md (GitHub profile and PR data).
Boundary markers: Largely absent in the instructions; the workflows rely on the model's grounding rather than explicit delimiters to distinguish instructions from analyzed content.
Capability inventory: The skill utilizes Bash, Edit, Write, Agent, and Skill tools extensively across all sub-pipelines.
Sanitization: No explicit sanitization or filtering of external content is mandated before interpolation into generated prompts or scripts.

workflow