How automation powers HafidWatch.
Nothing hidden.
Our articles are written by AI and reviewed by a person. This page explains exactly how, including what can go wrong.
What is automated, and what isn’t.
Most of the work at HafidWatch is done by software, including the writing. Here is the exact breakdown.
Automated
- Monitoring seven news sources every hour
- Scoring and selecting stories with a fixed formula
- Writing the article, headline, key points and “HafidWatch Take” (AI)
- Labeling sentiment and extracting tickers and names (AI)
- Checking each draft against our quality rules
- Publishing to the blog and six social channels
- Weekly digest compilation
Not automated
- Writing the editorial rules and the instructions the AI follows
- Choosing which sources to monitor
- Reviewing every article after publication
- Corrections and removals
- Advertiser, sponsor and affiliate relationships
Every story. 5 steps.
Every article goes through the following steps. The first four are automatic. The last one includes a person.
The RSS feeds of Cointelegraph, Decrypt, CNBC, Blockworks, Bloomberg, Protos and CryptoSlate are read every hour.
Each story receives a score from 0 to 1 from a fixed formula based on the outlet, the topic, how recent it is and the headline. No AI is involved in this step. Sponsored posts, digests, opinion columns and price-prediction pieces are filtered out, and the highest-scoring new story moves on.
An AI model from OpenAI (currently GPT-5) writes the article from the text of the original report, and assigns a sentiment label (bullish, bearish, neutral or mixed). Smaller OpenAI models (currently GPT-4.1) check for duplicate stories, extract names, and test whether the “HafidWatch Take” merely repeats the article.
A draft is published only if the original report had enough material, the draft reaches a minimum length without padding, and the AI itself grades the draft as fit to publish. Drafts that fail are discarded.
Drafts that pass are published automatically to the blog, Telegram, Discord, LinkedIn, Threads, Facebook and Bluesky. The founder then reviews every published article, corrects errors and removes articles that fall short.
What AI does, and what it doesn’t.
AI writes our articles. It does not do reporting, and it does not have the final word.
The AI is instructed to use only facts, figures and quotations found in the original report, and automated checks enforce part of that. AI models still make mistakes. They can misread a source, get a calculation wrong or state something more strongly than the evidence allows. That is why a person reviews every article, and why we correct errors publicly.
Every AI-written article says so in a line at the end, next to the name of the outlet whose report it is based on.
HafidWatch does not interview sources, obtain documents or break news. Each article works from one published report. Anything we add, such as a calculation or the “HafidWatch Take”, is marked as ours.
Every draft is written and checked under the same written rules. Consistent does not mean free of bias: AI models have blind spots of their own, and we change the rules when we find one.
The source list, the scoring formula, the AI instructions and the checks were written and are maintained by the founder, who also reviews every published article.
When something is wrong, a person corrects it and notes the correction on the article. AI does not issue corrections. We do.
The news pipeline sends the text of published news reports to OpenAI. It does not send information about readers or subscribers.
Transparency isn’t optional. It’s the product.
Automation is what lets a one-person publication cover markets every day. It is also the reason you should know how each article is made before you rely on it. If you find an error, tell us and we will fix it.
Questions about our AI automation process?
We’re happy to explain how any part of the pipeline works.