[maxbolgarin]

Summary

video preset: problems found

• Broken jump-to-moment links. The prompt asked for [HH:MM:SS](URL?t=Ns), but the link template is watch?v=ID&t=Ns. A model that followed the example wrote a second ?, which breaks the link.
• Wrong job. It called itself a summary ("Task: summarize"), while the bot labels it ▶ Analyze. There was no breakdown of the speaker's position, claims or arguments.
• Chunks split mid-argument. max_chunk_input_tokens: 35000 split anything past ~40 minutes into map-reduce fragments, so an argument was summarized in pieces.

What changed

• video v2 is an analysis. It covers:

  • the topic and the speaker's position;
  • claims, each with the arguments behind it and timestamps;
  • how the argument holds together (what rests on facts vs examples or emotion, unsupported claims, objections addressed or skipped), without judging truth, which is the fact-check's job;
  • numbers, quotes, and moments worth watching.

  The prompt now gives one explicit link example. The chunk cap is 90k, so a two-hour transcript (~50-60k tokens) fits in one call and the cap still stays under the low-TPM guard that test_video_preset_has_chunk_cap pins. The output budget is 8k.

• New video_summary preset (hidden). A retelling in order for someone who won't watch: what it's about, the video part by part with a timestamp per part, takeaways, and whether it's worth watching anyway.

• Surfaces.

  • Bot: the YouTube panel gets 📄 Summary (Y_SUM), which runs the existing single-item path with a preset override, like Fact-check.
  • CLI: the YouTube picker gets a Summary row (SUMMARY_SENTINEL).
  • --preset video_summary also works.

• Docs: docs/bot.md, docs/sources.md, presets/README.md.

Related issue

n/a

Test plan

• uv run pytest -q (2659 passed)
• uv run ruff check .
• uv run ruff format --check .
• tests/test_video_presets.py checks:
  • both presets exist in en and ru, are hidden, and open with a TL;DR;
  • links use the template's &t=, never URL?t=;
  • the chunk cap sits between 60k and 100k;
  • the analysis has a claims/arguments section and the summary is chronological;
  • the bot button and the CLI row run video_summary.
• Run both on a real video and read the output; the reports themselves weren't generated here.

Checklist

• PR title uses a Conventional Commits prefix
• Updated CLAUDE.md if architecture / invariants changed (n/a)
• Added or updated tests for the change
• No new network calls inside tests

🤖 Generated with Claude Code

https://claude.ai/code/session_019axffqQzzXeAQWXNrH2xpr

---

Generated by Claude Code

[github-actions]

🎉 This PR is included in version 1.9.0 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀