メニュー
Defensive protocols to prevent excessive usage allowance, quota, token, and context window burn-up.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Prevent and remove high-impact code-quality issues with principle-driven, risk-ranked recommendations and optional implementation.
Plan ytnova changes from behavior to implementation tasks with files, risks, invariants, and acceptance criteria. Use for design and technical-approach questions.
Implement approved ytnova tasks in C with minimal coherent edits, architectural safety, and terminal-verified proof.
Perform adversarial ytnova code review with severity-ranked findings, concrete fixes, and pass-fail gate decision.
Enforce strict ytnova bugfix loop: failing regression first, implementation second, passing proof third.
Run ytnova C quality gate for feature-sized changes using build, static analysis, valgrind, and pytest per docs/AUDIT.md.
| name | usage-economy |
| description | Defensive protocols to prevent excessive usage allowance, quota, token, and context window burn-up. |
You are operating under strict resource economy rules. Your system and user prompts consume usage allowance (also referred to as tokens, quota, computation, or context window) dynamically based on history depth and tool outputs.
You MUST follow these rules without exception to protect the user's usage allowance:
You MUST warn the user if they request an action that will consume an unnecessarily large amount of usage allowance. This applies broadly to ANY action that retrieves, generates, or parses massive amounts of text when a targeted approach exists.
Whenever you warn the user about a high-usage action, you MUST simultaneously provide the more economical alternative. You MUST always recommend the most targeted, precise, and lowest-context method available to achieve the user's goal.
If you have warned the user and suggested an alternative, but the user explicitly instructs you to proceed with the expensive action anyway, you MUST NOT prevent them. Execute their explicit command.
For the ytnova codebase, you MUST ALWAYS default to specialized MCP semantic tools (jCodeMunch or Serena) over generic bash tools to save usage allowance. You MUST use semantic discovery, AST node retrieval, and structural search instead of raw file reading, generic grepping, or unstructured directory listings.
You MUST paginate lists and searches unless explicitly instructed otherwise. If a conversation shifts to a completely different sub-system or topic, you MUST suggest the user opens a new chat session to drop the old context baggage.
