菜单
Test, tune, and regression-check LLM system prompts with a golden dataset. Use when the user wants to optimize a prompt, run prompt regression tests, manage `cases.yaml`, compare expected vs actual outputs, or follow a TDD workflow for prompts.
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
Open local review artifacts in Xiaoer Omia when users ask to open, preview, inspect, or review a file.
Run a measured improvement loop for a skill by defining a yes/no checklist, capturing a baseline, scoring candidate outputs, and logging each iteration's keep/revert decision. Use when a skill works inconsistently and you want evidence-driven iteration instead of random prompt edits.
Build a static HTML report site from CSV, JSON, Markdown, or HTML inputs. Use when the user wants to turn existing data into a human-readable site, generate a data showcase page, create a static report website, or optionally publish the built site.
Add offline multilingual support to static HTML, small web projects, or reports by extracting text into tokens, maintaining language dictionaries, applying translated outputs, and optionally embedding a local language switcher. Use when the user asks for bilingual support, tokenizing Chinese copy, offline language switching, or i18n for reports and lightweight sites.
Unified entrypoint for Obsidian note operations such as cleanup, merge, lint/fix, meeting archive, daily rollover, and docs sync. Use when the user wants one command surface over an existing notes-tools repo instead of calling many atomic scripts by hand.
Evaluate whether a proposed skill is worth building by scoring it against five moat-oriented questions: differentiation, vertical fit, learning loop, user difficulty, and defensibility. Use when someone proposes a new skill and you want a fast build / deepen / reject decision before implementation.
| name | prompt-tuner |
| description | Test, tune, and regression-check LLM system prompts with a golden dataset. Use when the user wants to optimize a prompt, run prompt regression tests, manage `cases.yaml`, compare expected vs actual outputs, or follow a TDD workflow for prompts. |
| version | 1.0.0 |
| author | Vivi |
| tags | ["prompt","llm","regression","tuning","tdd"] |
| platforms | ["all"] |
| dependencies | ["python3","requests","python-dotenv","pyyaml"] |
Treat prompts like code: prompt.txt is the source, cases.yaml is the golden dataset, tuner.py is the test runner, and CHANGELOG.md tracks prompt logic changes.
cases.yaml, prompt.txt, or a golden datasetprompt.txt and CHANGELOG.md after changes.Local check without API calls:
python3 tuner.py --check
Run full regression:
python3 tuner.py
Optional provider override:
python3 tuner.py --provider openai --model gpt-4o
prompt.txt, cases.yaml, .envMinimum validation command:
python3 tuner.py --check
This checks:
prompt.txt has a version headercases.yaml loads correctlyinput and expected--check for audits and review when API credentials are not configuredprompt.txt aligned with CHANGELOG.md