← All writing

Trying Codex and Grok Bot on My Daily News Briefing

Comparing my work briefing with a Grok Bot morning edition helped me see what makes a daily briefing worth opening.

I wanted to help our team keep up with AI and data science news, so I tried putting together a daily briefing with Codex and Grok Bot. I liked reading Grok Bot’s first result more than the briefings I’d been getting from Codex.

After those trials, I helped set up an automated pipeline at work to collect and summarise the latest news. We get a short briefing each day and can skim it to decide what to read in full.

Reading two briefings from the same day

On 17 November 2026, I read that day’s Grok Bot morning edition alongside our work briefing from 09:30 Hong Kong time. The work briefing had ten AI and data science stories in four groups. Grok covered thirteen stories across six categories, including Hong Kong news and K-pop, which I like to catch up on in the morning.

Both briefings picked up Hugging Face’s OpenRouter announcement. The work briefing gave me performance figures and routing settings I could use. Grok put the limits near the main point, spelling out that the rollout excluded image, video, and audio generation. I like being able to catch that while skimming.

They both covered TokenEconomizer, too. I paraphrased the two summaries in Cantonese.

Working briefing

2026-11-17

  • TokenEconomizer 據報喺 277 個修 bug 任務入面平均慳 18.4–30.9% 成本,長期用同一個 repo 嘅團隊可以試吓

Grok Bot

2026-11-17

  • TokenEconomizer 作者評估咗 277 個修 bug 任務,報告指平均每個任務嘅成本低咗 18.4–30.9%

The work briefing suggested trying TokenEconomizer if you work in the same repo regularly. Grok made it clear that the figures came from the paper’s authors. I like to know who’s reporting a result before I decide how much weight to give it. TokenEconomizer paper

What I asked Grok and Codex to do

I asked Grok for a short briefing in Traditional Chinese that reads like Hong Kong Cantonese. Alongside Hong Kong and international news, I wanted a quick look at US markets, updates from tech and AI companies, practical AI and data science, and K-pop. Two or three short sentences per story would give me enough to decide what to open. I capped each section at three stories and told Grok to skip it if it found nothing worth including.

I also asked Grok to read the original article or post, check its date and direct link, and skip the story if it couldn’t open the original or find a date.

For the setup I shared on GitHub, I asked Codex to pick and summarise stories using only the collector’s JSON. It had the titles and snippets, some of them cut short. Original briefing skill

I read through the prompt Grok drafted from my instructions. I liked that it spelled out how much context each story needed and when Grok should leave one out. That gave me a way to check the briefing against what I’d asked for. My 17 November task note reconstructs the main and backup instructions.

What Feedroom publishes on Threads

I helped prepare the internal briefing with enough technical detail for colleagues to use at work. For Threads, I wanted something readers could skim, so I kept it to one Cantonese headline per story. I publish it through Feedroom, a personal project I started once our team briefing was running.

I use a local schedule to start the Docker Compose job. Feedroom groups reports about the same event and sends source excerpts to OpenRouter to draft the headlines. It checks the draft before publishing. Feedroom on my personal GitHub

Feedroom’s local publishing flow. A host schedule starts the Docker Compose job, which collects and groups stories, validates OpenRouter drafts, stages PNG cards in private R2, and publishes to Threads. The publisher uses a SQLite journal to track dates, stories and post status.

I set the scheduled job to include image cards. Threads gets temporary links to fetch the PNGs from private R2. I keep the publishing history in SQLite so Feedroom can check whether it already posted that day before it tries again. Feedroom architecture

The 22 November edition was a six-card carousel. These are the first two headlines from its caption.

2026-11-22

  • Transformers 5.24 新增多模態模型 Inkling 2 Mini
  • AWS 推分散式 KV 快取技能 減自建 LLM 成本

Read the 22 November post on Threads

How I read each briefing

I read the team briefing for details I can use at work, like routing settings I might try. I use Grok’s morning page to catch up on a wider range of topics, from Hong Kong news to K-pop, and follow the source links for stories I want to read in full. On Threads, I publish one Cantonese headline per story through Feedroom, with cards readers can swipe through for more context. A quick read should tell me which stories are worth more of my time.