The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill defines a process to generate shell scripts (evaluate.sh and evaluate-checks.sh) used for automated benchmarking. These generated scripts include Python snippets for secure JSON encoding of dynamic content during LLM API interactions, demonstrating an expected use of code generation for build orchestration.
[INDIRECT_PROMPT_INJECTION]: As a factory for processing external domain materials, the skill has an inherent ingestion surface for untrusted data. It addresses this by providing explicit instructional guardrails in its best practices guide, mandating that fetched content be treated strictly as data rather than instructions.
Ingestion points: The input/ directory where users provide gold standards and study materials for research and evaluation.
Boundary markers: The references/skill-authoring-best-practices.md file contains a specific section on 'Untrusted third-party content' that instructs the agent to frame fetched data with explicit warnings to ignore any embedded commands.
Capability inventory: Capability to write files to the build workspace, generate shell scripts, and perform network requests via curl.
Sanitization: Implementation of instructional framing and the use of deterministic checks alongside LLM-as-judge evaluation to reduce reliance on potentially poisoned external inputs.
[COMMAND_EXECUTION]: The factory pipeline instructs the agent to execute the generated shell scripts to establish baselines and run validation checks during the drafting and autoresearch phases. These scripts perform standard operations using common utilities like grep, curl, and wc.
[EXTERNAL_DOWNLOADS]: The skill references external resources, including the mattpocock/skills repository for craft principles and the agent-skills-harness for reference implementations. These targets are well-known repositories within the developer community and are relevant to the skill's stated purpose of implementing industry best practices.

create-skill-autoresearch