all-in-distill

Run the distill script with --last or --video <ID> to print the full transcript and metadata, then have the LLM write the brief using the compact template. The script only loads transcripts; the LLM produces the brief to avoid templated output.

Use paste mode: save the transcript to a file and run distill.py --paste /path/to/file with optional --title and --date flags. This bypasses the SQLite database entirely and prints the transcript for LLM briefing.

Yes, use the --head flag (e.g., --head 4000) to truncate the transcript for previews and spot-checks. Feed the full 80–100k character text only when generating a real brief.

The brief follows a compact Telegram template: date and video ID header, one-line core thesis, 3–5 key takeaways with numbers, company positioning with bullish/neutral/bearish calls, and one portfolio watch sentence, kept under ~2000 characters.

A previous design shipped a hardcoded brief identical for every episode, which was a regression. The script deliberately only loads transcript content so the LLM writes briefs tied to each episode's actual content.