Daily market-intel scan
A morning research brief and a weekly decision memo.
What it is
The intel scan is a research agent on a schedule. Every morning at 08:03 it pulls fresh items from fixed sources and scores up to three product ideas on buyer demand, revenue math and hype. Every Monday at 09:13 a second job compares the week's notes, checks competitors' official pricing pages and returns one verdict: GO, WATCH or PASS. Each run writes a cited note and sends a short brief to Telegram. The agent is allowed to report that nothing is worth pursuing.
By the numbers
What it means for your team
A team can point the same setup at its own sources, such as competitor pricing pages, review sites and industry news, scored against its own written rules. Each morning brings a short cited brief to a phone, and the agent can say there is nothing worth acting on.
Status
No idea has cleared the weekly bar so far, and no buyer test has a recorded result. The first version failed often and was paused in July. The rebuilt version still lost runs to an expired sign-in from 2026-08-28 to 09-03 and an exhausted quota from 09-20 to 09-25. Both jobs moved to Claude Sonnet 5.5 on 2026-10-08; that morning's run fell back to GPT-5.6 Sol after a lapsed sign-in. It has not yet been run for a client.
Fetch before reasoning
Before the model runs, a 139-line Python script collects 15 Hacker News top stories, 8 Show HN posts, 25 Reddit entries from three startup subreddits and the current Google Trends feed. When a local Chrome session is available, it adds X search results.
Blocked sources on record
The script keeps a list of sources that failed before, such as Reddit's JSON API and Product Hunt, and the agent does not retry them.
Morning rubric
Demand, revenue and inverse hype are each scored from 1 to 10. A demand score above 4 needs a named buyer who already pays for something similar. Posts on X count only toward hype. The agent may not invent prices, traffic or conversion rates.
Weekly rubric
Demand and revenue are weighted 25% each, competition 20%, durability and use of existing assets 15% each. Anything under 7.5 out of 10 returns WATCH or PASS, and every finalist needs a 72-hour buyer test with a numeric target and a stop rule.
Records
Each run writes a note with fields for verdict, buyer and sources, saves a snapshot of its sources and logs incidents to SQLite. From June to early July, a static dashboard indexed the research files; it is now paused.