The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data from external sources, including scientific article URLs, GitHub issues, and free-text specifications provided by the user (described in § 2 of SKILL.md and references/record_outcome.md). This external content can potentially contain malicious instructions aimed at influencing the agent's behavior during the ML iteration loop.
Ingestion points: GitHub issue URLs, scientific article URLs, and user free-text (found in SKILL.md § 2 and references/record_outcome.md).
Boundary markers: The skill employs several safety gates including G-DESIGN (explicit user approval of design notes before script creation) and G-RUN (explicit user approval before execution), as well as a mandatory pre-flight checklist.
Capability inventory: The skill possesses the capability to execute shell commands via pixi run python for experiment scripts and audits, and perform network operations using gh issue comment (found in SKILL.md § 3 and references/record_outcome.md § 6).
Sanitization: The provided instructions do not specify technical sanitization or filtering of the external content, relying primarily on human-in-the-loop approval gates.
[COMMAND_EXECUTION]: The skill uses the pixi package manager to execute generated Python scripts for Machine Learning experiments (pixi run python experiments/NN_<short_name>.py) and audit diagnostics (audit/NN_<short_name>.py). It also uses the GitHub CLI (gh issue comment) to post result summaries to external issues. These commands are executed based on the agent's generation and user approval.
[DYNAMIC_EXECUTION]: The skill generates and executes Python scripts (experiments/NN_*.py, audit/NN_*.py, and scratch/*.py) based on designs synthesized from research papers or GitHub issues. While these scripts are created locally, they are executed at runtime, which is a standard requirement for the skill's primary purpose of ML iteration.

iterate-ml-experiment