convez.ai
SkillDados & AnalyticsAvançado
liaoliang1981

Dietary Supplement Scorer

liaoliang1981/dietarysupplement-scorer

Pontuação determinística de suplementos com arquitetura multi-camadas

Ver no GitHub
InicialPluginScore · de 0 a 100

Engine determinístico para pontuação de suplementos alimentares com fonte única de verdade. Avalia fórmula, qualidade e segurança através de Claude Skill, ChatGPT e web, mantendo algoritmo sincronizado em todas as camadas.

Instalação

$npx convez-ai liaoliang1981/dietarysupplement-scorer

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
v0.1.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Validação automática de fórmulas de suplementos nutricionais
  • Avaliação comparativa de qualidade e segurança entre marcas
  • Integração em plataformas de saúde e e-commerce

Veredito

Projeto sólido com arquitetura bem pensada (engine determinístico + três camadas de integração). Excelente para quem precisa de scoring reproduzível e auditável; requer ajuste dos valores de referência por especialista antes de uso em produção.

README

A tradução para o português deste README ainda está na fila.

dietarysupplement.ai supplement scorer — monorepo

English · 简体中文


A deterministic supplement scoring capability, delivered through three calling layers: Claude Skill, ChatGPT, and a browser-based agent. All categories, brand-agnostic, global. Scoring runs a deterministic engine (not LLM-estimated); models only handle conversation and explanation.

Single source of truth (the key idea)

Only edit engine/, then run python build.py to sync to all layers. Do not hand-edit any "generated" file.

engine/
├── data.json          ← single data source: reference tables + weights (edit here)
├── score.py           ← single algorithm (edit here)
└── test_engine.py     ← assertion tests
build.py               ← sync tool: pushes engine → the three layers below
claude-skill/
└── supplement-scorer/ ← Claude Skill (SKILL.md + generated scripts/score.py + references/)
chatgpt/
├── score.py           ← generated: self-contained (data inlined, for Code Interpreter)
├── gpt-instructions.md, openapi.yaml, server.py, dev-doc.md
web/
└── index.html         ← browser-based agent (CFG + REFERENCE data block generated by build.py)

After editing:

python build.py                  # sync all four locations
python engine/test_engine.py     # assertion tests (7/7)

Verification: the same "premium iron supplement" scores 97 across engine tests, chatgpt/score.py, claude-skill, and web. Change weights or dose ranges in engine/data.json, build, and all four update together.

How to use each layer

  • Claude Skill — package claude-skill/supplement-scorer/ and upload to Claude.ai (Settings → Capabilities → Skills), or drop it into the Claude Code skills directory.
  • ChatGPT — follow chatgpt/dev-doc.md: Custom GPT uploads chatgpt/score.py + enable Code Interpreter + paste gpt-instructions.md; for stability, use server.py + openapi.yaml as an Action; for your own product, use the Responses API.
  • Web — web/index.html is an interactive browser-based agent (runs directly in the Claude.ai artifact environment; standalone deployment requires a backend proxy to hold the model API key).

Caveats

  • The reference values in engine/data.json are a v0 international baseline, manually entered, non-authoritative. National RDAs/ULs and permitted claims vary; before going live, a domain expert must review and tag each value with a source per market.
  • The safety-deduction constants (25 / 15) are hardcoded in the engine algorithm (algorithm constants, not data).
  • Web is JS, the rest is Python: build.py syncs data (from data.json); the algorithm itself is maintained on both sides as a mirror. When changing algorithm logic, update both.

Design dialogue

Why this architecture (deterministic engine vs LLM scoring, fact / weight separation, market × category × vertical scope discipline) is documented in:

📄 docs/design-dialogue-2026-06-08.md — 24-question design conversation, EN + 中文.

License

MIT


一套确定性补剂打分能力,通过三个调用层交付:Claude Skill、ChatGPT、浏览器内 agent。 全品类、不分品牌、面向全球。打分跑确定性引擎(不靠模型猜),模型只负责对话与解释。

单一真相源(重点)

只改 engine/,然后 python build.py 同步到所有层。 别手改任何"生成的"文件。

engine/
├── data.json          ← 唯一数据源:参考表 + 权重(改这里)
├── score.py           ← 唯一算法(改这里)
└── test_engine.py     ← 断言测试
build.py               ← 同步器:把 engine 推到下面三层
claude-skill/
└── supplement-scorer/ ← Claude Skill(SKILL.md + 生成的 scripts/score.py + references/)
chatgpt/
├── score.py           ← 生成:自包含(数据内嵌,供 Code Interpreter)
├── gpt-instructions.md, openapi.yaml, server.py, dev-doc.md
web/
└── index.html         ← 浏览器 agent(CFG+REFERENCE 数据块由 build.py 生成)

改完跑:

python build.py          # 同步四处
python engine/test_engine.py   # 断言测试(7/7)

校验:同一份"优质铁剂"在 engine 测试、chatgpt/score.py、claude-skill、web 四处都得 97。改了 engine/data.json 的权重或剂量区间,build 后四处一起变。

各层怎么用

  • Claude Skill —— 打包 claude-skill/supplement-scorer/ 上传到 Claude.ai(Settings → Capabilities → Skills),或放进 Claude Code 的 skills 目录。
  • ChatGPT —— 按 chatgpt/dev-doc.md:Custom GPT 传 chatgpt/score.py + 开 Code Interpreter + 粘 gpt-instructions.md;要稳就用 server.py + openapi.yaml 做 Action;接自家产品用 Responses API。
  • Web —— web/index.html 是浏览器内的交互式 agent(在 Claude.ai artifact 环境直接跑;独立部署需后端代理保管模型 API key)。

注意

  • engine/data.json 的参考值是 v0 国际基线、手工录入、非权威;各国 RDA/UL 与可声称项不同,上线前需专业人士复核并按市场挂来源。
  • 安全扣分常数(25/15)在引擎算法里写死(算法常数,非数据)。
  • Web 是 JS、其它是 Python:build.py 同步的是数据(从 data.json);两边的算法各自维护、保持镜像。改算法逻辑时记得两处都改。

设计对话

为什么用这套架构(确定性引擎 vs LLM 估分、事实/权重分离、市场 × 品类 × 垂类 的范围纪律)的完整推理过程:

📄 docs/design-dialogue-2026-06-08.md —— 24 轮设计对话,中英双语。

License

MIT

Ver o original no GitHub

Bom para

Relacionados