Idea Probe

I did not want an “AI summariser” — a summary is read once and gone by the next morning. What I wanted was to break other people’s arguments into verifiable minimum units, then compare them against an existing archive: what is genuinely new, and what keeps coming up again and again.

Idea Probe is that. Input: Reddit threads. Output: structured claim + evidence entries archived into a format-compatible archive.jsonl.

The flow

1. Fetch Reddit hot threads (needs REDDIT_CLIENT_ID / SECRET, via proxy)
2. Use DeepSeek to distil each thread into claim + evidence structure
3. Archive into the existing archive.jsonl format, ready for later retrieval and comparison

Engineering notes

• PRD before code, and the PRD has a dedicated “reality calibration” section that separates what is aspirational from what this version can genuinely do
• A --check health command: one command verifies API key connectivity, proxy reachability and the Reddit channel, and prints a report — for a script depending on several external services, locating the failure usually costs more time than writing the feature
• Honest status: it is an MVP today, offline self-tests pass, and the remaining to-dos are filling in keys, wiring up comparison against the existing archive format, and installing the scheduler

Why it is worth building

The bottleneck in information gathering stopped being “seeing it” a long time ago — it is remembering it and comparing it. The point of this project is to turn reading into an accumulating structured asset instead of a feed that is refreshed and forgotten.