Small decisions. Interesting possibilities.Submit contentSubmit
WeFlow CLI preview

WeFlow CLI

For each article, the client sends a state containing the title and up to 4,000 characters of body.

Added to Jevfast

How it uses Jev

For each article, the client sends a state containing the title and up to 4,000 characters of body. One request asks for its topic (Choice), practical relevance to the stated environmental-science/computing reader (Score), and whether it contains immediately usable material (Noul). The caller stores these structured answers; summaries and other generated prose remain in the LLM path.

What Jev decides

Illustrative source-derived request: classify one articleExample answers · not a recorded Jev response · Source ↗
Question 1 · topic
YOUR APP
INSTRUCTION

Which topic does this article belong to?

STATE

Article title and the first 4,000 body characters, formatted as “标题:… 正文: …”. The state does not include date, source, or prior labels.

JEV · CHOICE
  1. AI
  2. 投资
  3. 新闻
  4. 文学
  5. 学术
  6. 政治
Question 2 · worth including
YOUR APP
INSTRUCTION

Does the article contain something concrete the reader can use today, such as a new tool, method, data source, codebase, or reproducible result, rather than only informational news, opinion, or personal writing?

STATE

Article title and the first 4,000 body characters, formatted as “标题:… 正文: …”. The state does not include date, source, or prior labels.

JEV · NOUL
YesNo

App workflow

  1. Run the daily article workflow

    The CLI fetches articles for a requested date and calls the Jev classification path for each article when configured.

  2. Classify articles in parallel

    Up to six worker slots call the article-level client. Successful decisions are attached to article records; a failed Jev classification for one article falls back to the full LLM classification prompt.

  3. Generate article text

    For successfully judged articles, the LLM receives the summary-only prompt; Jev does not generate summaries.

Why it is interesting

It separates topic, usefulness, and inclusion into distinct typed decisions, then lets the daily workflow fall back to its previous LLM classification path when a Jev call fails.