Look Tongji Notes

Overview

Use the bundled CLI to:

1. configure Tongji IAM credentials (saved in this skill folder only),
2. list recent courses after login,
3. transcribe a chosen lecture to SRT + TXT,
4. download lecture slide snapshots to local files,
5. write a Markdown study note (by the current agent) using transcript + slide images.

This skill is intentionally CLI-first (no frontend).

Quick start (workflow)

When you see these trigger phrases, follow this mapping:

• look-tongji:setup -> run python "<SKILL_DIR>/scripts/look_tongji.py" setup
• look-tongji:list -> run python "<SKILL_DIR>/scripts/look_tongji.py" list
• look-tongji:slide -> run python "<SKILL_DIR>/scripts/look_tongji.py" slide ...
• look-tongji:transcribe -> run python "<SKILL_DIR>/scripts/look_tongji.py" transcribe ... (aliases: transcript, trans)
• look-tongji:note -> run python "<SKILL_DIR>/scripts/look_tongji.py" note ... (default: transcript + slide in parallel)

Where:

• <SKILL_DIR> is the directory that contains this SKILL.md.

Output + storage conventions

• Credentials: saved to <SKILL_DIR>/.env (and ignored by <SKILL_DIR>/.gitignore).
• Auth cache: saved to <SKILL_DIR>/state/ (JWT token cache and last selection).
• Artifacts (subtitles + transcript + slides + notes): saved to the agent working directory in ./tongji-output/ (relative to where the command is executed).

Never ask the user to paste passwords into chat. Prefer interactive terminal input (or environment variables) and keep .env out of version control.

References

When you need more details, read:

• references/url-guide.md for how to obtain and parse lecture URLs
• references/troubleshooting.md for common setup/login/ASR issues
• references/security.md for credential/cache/output safety notes

Command: look-tongji:setup

Goal: verify basic dependencies and store credentials in <SKILL_DIR>/.env.

Run:

python "<SKILL_DIR>/scripts/look_tongji.py" setup

Non-interactive options (avoid passing passwords via CLI when possible):

python "<SKILL_DIR>/scripts/look_tongji.py" setup --username "<YOUR_ID>" --password "<YOUR_PASSWORD>" --overwrite

If dependencies are missing, install them from <SKILL_DIR>/requirements.txt.

Command: look-tongji:list

Goal: login and list recent courses with indices.

Run:

python "<SKILL_DIR>/scripts/look_tongji.py" list

If you need to search a course by name (and it may not appear in "recent"), use the full course list method TongjiClient.get_all_courses() for higher accuracy.

CLI shortcut:

python "<SKILL_DIR>/scripts/look_tongji.py" list --all --query "<COURSE_NAME_KEYWORD>"

If you already know which course index to choose (for non-interactive runs):

python "<SKILL_DIR>/scripts/look_tongji.py" list --choose 1

This saves the selected course to <SKILL_DIR>/state/last_course.json.

Command: look-tongji:note

Goal: default workflow for one lecture: transcribe + download slides + write notes.

Preferred inputs (ask user for one of these):

• Lecture page URL from look.tongji.edu.cn (best effort parsing for course_id + sub_id)
• course_id + sub_id (most reliable)

Transcript-only command with a lecture URL:

python "<SKILL_DIR>/scripts/look_tongji.py" transcribe --lecture-url "<LECTURE_URL>"

Transcript-only command with IDs:

python "<SKILL_DIR>/scripts/look_tongji.py" transcribe --course-id "<COURSE_ID>" --sub-id "<SUB_ID>"

If only course_id is provided, the CLI can list the latest lectures and let you choose:

python "<SKILL_DIR>/scripts/look_tongji.py" transcribe --course-id "<COURSE_ID>"

Default agent behavior for look-tongji:note

Unless the user explicitly says "不要下载PPT/slide" (or equivalent):

1. Run note command for the same course_id + sub_id (internally does transcript + slide in parallel).
2. Unless the user explicitly says "不要大纲/不要时间线/no outline/no timeline" (or equivalent), generate a timeline outline from the SRT subtitles and write:
   • <course_id>_<sub_id>_timeline.txt in the same tongji-output/ folder.
3. After both finish, write notes using both transcript and slide images.
4. If slide download fails but transcript succeeds, still produce notes and mention slide failure briefly.

Default combined command:

python "<SKILL_DIR>/scripts/look_tongji.py" note --course-id "<COURSE_ID>" --sub-id "<SUB_ID>"

If user explicitly requests transcript only:

python "<SKILL_DIR>/scripts/look_tongji.py" note --course-id "<COURSE_ID>" --sub-id "<SUB_ID>" --no-slide
# or:
python "<SKILL_DIR>/scripts/look_tongji.py" transcribe --course-id "<COURSE_ID>" --sub-id "<SUB_ID>"

Artifacts produced by the CLI

The CLI writes to ./tongji-output/ (relative to the current working directory):

• <course_id>_<sub_id>.srt
• <course_id>_<sub_id>.txt
• <course_id>_<sub_id>.json (metadata)
• slide_<course_id>_<sub_id>/ (slide images + index.json)

Additional artifact written by the agent (timeline outline)

By default (unless the user asks not to generate an outline), after the CLI writes SRT, generate a timeline outline file for quick video overview:

• <course_id>_<sub_id>_timeline.txt

Stabilize timeline input before generation:

• If the SRT is too long to fit in context, run python "<SKILL_DIR>/scripts/timeline_tools.py" srt-sample --srt "<SRT_PATH>" --out "<SAMPLED_SRT_PATH>" --max-chars 15000 and only use the sampled output as the timeline context.

Format requirement (strict): each line must be:

• 起始时间-结束时间：课程阶段内容

Use the following timeline prompt (output must be in Simplified Chinese; output only the timeline lines):

You are a course TA. Your job is to turn the provided SRT subtitles into a concise video timeline outline for quick overview.

CRITICAL OUTPUT RULES:
1) Output ONLY the timeline lines. No title, no explanations, no Markdown, no numbering, no blank lines.
2) One line per segment, STRICTLY in this format (use the full-width Chinese colon '：'):
   MM:SS-MM:SS：<课程阶段内容（简体中文）>
3) Time format is MM:SS (minutes may exceed 59; seconds are always 2 digits; minutes are at least 2 digits and can be 3+ digits when needed). Examples:
   00:00-05:30：课程定位与考核说明
   05:30-12:00：电路理论学科体系介绍
   92:45-98:10：案例推导与易错点总结
4) The timeline MUST start at 00:00. Segments must be strictly increasing AND contiguous:
   - Previous end time == next start time
   - No overlaps, no gaps
5) Cover the whole lecture. Let the last segment end at (or very close to) the last subtitle end time.

CONTENT RULES:
6) Segment summaries must be in Simplified Chinese short phrases, using course terminology.
7) Fix obvious ASR typos/homophones when the meaning is clear, but DO NOT fabricate content not present in subtitles.
8) If a segment includes important admin items (assignment / exam / attendance / grouping / bonus points / deadlines), explicitly mention it in that segment.

GRANULARITY:
9) Prefer 10–20 segments for a typical 60–120 minute lecture; each segment usually 3–10 minutes.
10) Make segments shorter only when there is a clear topic shift or a key announcement.

Now generate the timeline outline from the SRT below:

SRT handling note (when context is limited):

• If the SRT is too long to fit in context, sample subtitle blocks uniformly across the whole file (do not only take the beginning), then generate the outline.

Post-processing / strict validation:

• After the agent produces timeline text, validate and normalize it using:
  • python "<SKILL_DIR>/scripts/timeline_tools.py" timeline-normalize --input "<TIMELINE_TXT>" --srt "<SRT_PATH>" --tolerance 1
• This enforces: starts at 00:00, segments are contiguous (no overlaps/gaps), and the last segment end is close to the last subtitle end time.

Important transcript-reading rule:

• For note writing, do not use <course_id>_<sub_id>.json as the main transcr