菜单
Raise recall in code review — get Claude to report every bug it finds, including low-confidence ones. Use for code review harnesses, security audits, or any workflow where missing a real bug is worse than a few false positives.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Make Claude default to implementing changes rather than just suggesting them. Use when the model keeps proposing edits instead of applying them, or when you want a bias toward action in an agentic coding context.
Stop Claude from leaving behind temporary scripts, scratch files, and iteration helpers after a task. Use when you want a clean workspace — no extra files beyond the requested deliverables.
Make Claude's responses shorter and more focused. Use when the model is over-explaining, adding non-essential context, or producing verbose answers for simple lookups.
Make Claude ask before taking destructive or hard-to-reverse actions (rm, force-push, dropping tables, posting publicly). Use in autonomous agents that touch shared systems or in any context where the blast radius of a wrong action is high.
Structured, hypothesis-driven research across large corpora. Use for complex research tasks, literature reviews, competitive analysis, or anywhere Claude needs to synthesize information from many sources with tracked confidence.
Get visual variety by having Claude propose multiple design directions before building. Use for open-ended design briefs, landing pages, brand work, or any frontend where you want to pick a direction rather than accept the default.
| name | vojta-bug-hunt |
| description | Raise recall in code review — get Claude to report every bug it finds, including low-confidence ones. Use for code review harnesses, security audits, or any workflow where missing a real bug is worse than a few false positives. |
Opus 4.7 is a stronger bug-finder than prior models, but phrases like "only high-severity," "be conservative," or "don't nitpick" make it self-filter findings it investigated. Separate finding from filtering.
Report every issue you find, including ones you are uncertain about or consider low-severity. Do not filter for importance or confidence at this stage - a separate verification step will do that. Your goal here is coverage: it is better to surface a finding that later gets filtered out than to silently drop a real bug. For each finding, include your confidence level and an estimated severity so a downstream filter can rank them.
Be concrete about the bar, not qualitative:
Report any bugs that could cause incorrect behavior, a test failure, or a misleading result; only omit nits like pure style or naming preferences.
If you have a downstream verification, dedup, or ranking stage, explicitly tell the model its job at the finding stage is coverage rather than filtering. Iterate against a held-out eval set to measure recall/F1.