LearnPrompt/ai-news-radar: A 24-Hour AI News Radar You Can Fork

The problem: too many sources, one 24-hour window

AI News Radar exists because the same announcement arrives from an official blog, a changelog, a post on X and three aggregators, and the reader still has to open dozens of pages to work out which version is worth reading. The README frames the project as a response to that repetition: the author describes opening dozens of pages, filtering duplicates by eye and guessing which item matters.

The project is for two audiences. The first is a reader who wants a page showing the last 24 hours of AI, model, product, developer and research updates, grouped by day and split into a curated pool and a broader pool. The second is a developer or Agent user who wants to own the source list. The README describes a three-step path: let an Agent read the news for you, read the site yourself, then fork the repository and run your own filter.

What separates it from a feed reader is the ordering of work. The project states that it judges source quality before it fetches, rather than fetching everything and sorting afterwards. That order is the whole editorial argument, and it is also the reason the repository ships a Skill for evaluating sources rather than only a scraper.

How the pipeline turns sources into data/*.json

The README's flow diagram shows a source list going through a classification step, then splitting into five fetch paths: official RSS or changelog, personal OPML or RSS, public GitHub feeds or JSON, public pages with a Jina fallback, and AgentMail subscriptions. A sixth branch skips high-risk sources entirely. Everything that is fetched then passes through deduplication and normalisation, AI relevance scoring and tagging, story merging with multi-source evidence, and source health accounting.

The output is a set of static JSON files, and the front end reads only those files. There is no backend service. GitHub Pages is described as the canonical source of the data, with the Vercel site presenting the same files. The documented files include data/daily-brief.json for the curated 20-item brief, data/top3-personas.json for the three-persona commentary on the day's top three stories, data/latest-24h.json for strongly AI-related messages, data/latest-24h-all.json for the broader pool with score >= 0.3, data/latest-24h-all-raw.json for an unfiltered dev-only dump, data/source-status.json for fetch status and source health, data/stories-merged.json for merged events, and data/merge-log.json for the merge process.

Two behaviours are worth noting because they affect what you see. If daily-brief.json is missing, the README says the page falls back to a candidate signal list. If stories-merged.json exists, the page uses the full story pool to fill in later story lines. The front end also accepts a ?data=<directory> parameter, which points it at a different data directory and remembers the choice locally, so you can preview a branch or pull request without editing code.

Installing the ai-radar Skill and getting a first brief

The README's fastest path is the Skill, installed with npx. The project states that this path needs no API key, no login and no server.

npx skills add LearnPrompt/ai-news-radar -s ai-radar -g

After installation, the README says to ask the Agent a single question. The example it gives is:

今天AI圈有什么？

The expected result is a daily brief drawn from the generated data, with the persona commentary referenced from data/daily-brief.json and data/top3-personas.json. If you prefer the browser, the project points at news.learnprompt.pro, with learnprompt.github.io/ai-news-radar listed as a data source and backup. The site defaults to a mobile view; a view switch in the top right moves to the classic desktop interface under /classic/, and the URL parameters ?view=mobile, ?view=classic and ?view=auto select a view directly. Both views read the same data directory.

The second path is to fork the repository and let GitHub Actions generate the JSON. The README's fork guide lists five steps and the repository contains .github/ plus a workflow referenced as update-news.yml. The Python dependencies for the pipeline are pinned in requirements.txt: requests 2.32.3, beautifulsoup4 4.12.3, feedparser 6.0.11 and python-dateutil 2.9.0.post0. Advanced sources are meant to be supplied through GitHub Secrets or local environment variables, and examples/advanced-sources.env.example is the file the repository provides for that.

Personas are markdown files, and that is the real configuration surface

The three commentary voices, pragmatic, cynic and paper-police, are not code. Each one is a markdown file in personas/ with frontmatter and a system prompt. Changing the tone means editing one file; adding a voice means writing a new file in the documented format and opening a pull request. The README describes pragmatic as the default, focused on what is useful to a developer today, cynic as attacking marketing language while staying factual, and paper-police as accepting only papers, code and benchmarks.

This is a more honest configuration model than a settings page, because the prompt is the artefact and it lives in version control. It also means the quality of the commentary is bounded by the model behind it. The README states that LLM commentary requires DEEPSEEK_API_KEY on the upstream side. Without it, the pipeline still runs end to end and degrades to rule-based scores, and the README says both the page and the Skill keep working.

The same key gates the title rewriting feature introduced in v0.9, which sends short or jargon-heavy titles to an LLM with surrounding context, falling back to r.jina.ai when the project's own fetch fails. The README is explicit that without the key the original title is kept and nothing else in the flow changes. One more detail from the v0.9 notes: the one-line recommendation reason on curated cards is generated by the pipeline, and when no real reason exists the front end hides the block instead of filling it with a template sentence. That is a deliberate choice to show less rather than show filler.

Where the design shows strain

The most visible limitation is that the three-persona web display has been taken offline. The README's v0.8 note says the page presentation is archived pending a style redesign, while the data pipeline continues to generate persona scores and commentary every day and the Skill brief is unaffected. So the feature that gives the project its name in the README headline, three-flavour commentary, currently reaches you through the Skill and the JSON files rather than through the site.

The v0.9 notes also record a correction rather than a feature. The aggregator source zeli, a Hacker News 24-hour hot list, had been whitelisted wholesale and now goes through the same AI relevance scoring as other aggregators. The same release added validation for bilingual title translation so that refusal text and degenerate output are detected and the original title is restored. Both entries describe a pipeline that was previously letting low-quality items through, which is a useful signal about how much of the filtering is heuristic.

There is also a dependency question. The pipeline is designed around GitHub Actions generating static JSON, so the refresh cadence is tied to that workflow. The README does not document rollback for a bad data run, and the legacy/ directory holding the old three-view screenshots is described as kept only until mid-August 2026. Anyone forking should expect to read scripts/ and .github/ rather than rely on documentation alone.

Alternatives and the difference in approach

The README positions AI News Radar alongside two sibling sites, AI MAP for a seven-day heat map refreshed every 12 hours and GoodCase for use cases and prompts. Those are complements, not substitutes, and the README describes them as one pipeline: discover, track, use.

The closer comparison is a general feed reader or RSS aggregator. A reader like that keeps every item and leaves ranking to you; AI News Radar scores items for AI relevance, merges multi-source coverage of one event into a single card with a multi-source label, and publishes a curated pool separately from the broader pool. The difference that matters in daily use is the merge step. A feed reader would show you five headlines about the same release; this pipeline folds them into one card and lets you expand the individual titles.

The trade-off runs the other way too. A feed reader is indifferent to subject matter and will happily track anything you subscribe to. AI News Radar is built around AI relevance scoring, so a source whose AI content is a small fraction of its output will be scored down, which is the intended behaviour but also means the tool is the wrong choice for tracking a non-AI beat.

Licence and the cost of keeping a fork current

The repository is MIT licensed, which permits reuse and modification provided the licence and copyright notice are retained. That is a permissive starting point for a fork, but it says nothing about the data you pull in or the terms of the sources you add. If you connect private OPML files, mailbox content or paid APIs, the README's guidance is to keep those out of the repository and supply them through GitHub Secrets or local environment variables; examples/advanced-sources.env.example is the template. That guidance is about avoiding committed secrets, and it is worth following literally.

Upgrade cost depends on how far you diverge. If you only edit personas/ and swap the source list, upstream changes to the pipeline and the front end arrive as normal merges. If you rewrite the scoring or the merge logic, you own that divergence. The release history shows the interface changing shape between v0.7 and v0.9, with the three-view layout collapsed into a single-layer architecture, so a fork that customises the front end should expect to reapply those changes. The last push to the repository was on 2026-07-14, so check the commit history before assuming a fix is coming.

Editorial conclusion

Adopt it if you want a daily AI brief you can inspect and fork, or if you are willing to run GitHub Actions and edit personas/*.md to change the editorial voice. Do not adopt it if you need a hosted service with an SLA, or if you cannot accept that the pipeline depends on GitHub Actions and DeepSeek for its LLM features. Before forking, check that the repository still receives pushes (the last one was on 2026-07-14), confirm which data/*.json files your branch actually generates, and verify whether your fork inherits the DEEPSEEK_API_KEY secret, because without it the pipeline falls back to rule-based scores and skips title rewriting.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes