Anibon Timestamper (Local LLM Edition)

Overview & Triggers

Optimized for local LLMs with limited context windows (Gemma 4, Qwen 2.5, etc.). Sequential chunk loop — no parallel subagents, no cloud.

🧠 Goldfish Brain Rules (CRITICAL)

Violating the letter of these rules is violating the spirit of these rules.

Red Flags — STOP and Call a Tool

If you catch yourself doing any of the following, STOP GENERATING TEXT AND CALL A TOOL:

• Generating "Wait", "Actually", or "Hold on" (infinite reasoning loop).
• Thinking "I will do both chunks now to be efficient."
• Thinking "I don't remember the prompt format, I'll just guess."
• Running ls to check if a file exists.
• Outputting generated timestamps directly into chat instead of calling the write tool.

Anti-Rationalization Table

Core Constraints

1. One tool per turn: Never batch tool calls across chunks.
2. Process inline: Read .txt, write timestamps yourself. No subagents.
3. No <think> tags: Never wrap reasoning in <think>.
4. Handoff trigger: User says "handoff", "stuck", or "context > 10%" → IMMEDIATELY write state file and halt.

🗺️ Plugin Directory Map (Do NOT use ls)

You already know where everything is. Resolve [SKILL_ROOT] in Step 0.

• Scripts: [SKILLROOT]/scripts/preparevideo.py
• DB + Check Scripts: [SKILL_ROOT]/scripts/

- fetchfgodb.py, fetchygodb.py, checksections.py, packtimestamps.py

• Workspace: [WORKSPACE] — set in Step 0

- [WORKSPACE]/chunks/chunkXX.txt - [WORKSPACE]/chunkoutputs/chunkXXoutput.md - [WORKSPACE]/parts.json - [WORKSPACE]/anibontimestamps.md - [WORKSPACE]/anibontimestamper_state.json

🧭 Steps

[!IMPORTANT]
REQUIRED SUB-SKILL (FIRST): Run preparing-tools to verify yt-dlp, python3, sqlite3. Do NOT proceed if any tool is missing.

Step 0: Resolve Plugin Root & Workspace (Cross-Platform)

Find [SKILL_ROOT]: Look at the <skill location="..."> tag at the top of your prompt.

• Strip the filename SKILL.md (or with backslashes \ on Windows).
• Replace all \ with /. The result is [SKILL_ROOT].

🚨 ANTI-TYPO: Plugin repo = anibon-stream-synthesis (HYPHENS). Skill folder = anibon-timestamper (HYPHENS). NEVER use underscores. Copy paths directly; do not retype from memory.

Verify Python:

Mac/Linux:

uname -a && python3 --version

Windows (PowerShell):

python --version

Set [WORKSPACE]: Unless user specifies a path, default to:

• Mac/Linux: ~/youtube<videoid>_workspace
• Windows: Use forward slashes — C:/Users/<username>/youtube<videoid>_workspace

Step 1: Initialization

1. Ask output language: "What language for the timestamps? (e.g. Thai, English)" — do NOT proceed to timestamp generation until confirmed.
2. Channel ownership check: Run:

``bash yt-dlp --print uploader,uploaddate "VIDEOURL" ` Verify uploader matches one of: Phuboat, ปู่โบ๊ต, โบ๊ต, Boat, ANIBON`. If no match — do NOT call the speaker "Boat".

1. Royal/political content: If transcript mentions Thai royalty, royal succession, or sensitive political topics → use metaphor-based masking.

REQUIRED SUB-SKILL: masking-royal-news

Step 2: Download & Chunk

Mac/Linux:

python3 "[SKILL_ROOT]/scripts/prepare_video.py" "VIDEO_URL" --format txt --block 300 --overlap 30

Windows (PowerShell):

python "[SKILL_ROOT]/scripts/prepare_video.py" "VIDEO_URL" --format txt --block 300 --overlap 30

Local LLM Note: Always use --format txt. Do NOT use --vision — local models cannot process images.

(If blocked by YouTube, ask user for cookies file or rawtranscript.json. If subtitles/captions are completely missing, transcribe the audio locally using whisper.cpp as detailed in [BUILDWHISPERCPPGUILD.md](../anibon-timestamper/references/BUILDWHISPERCPP_GUILD.md).)

Step 3: Sequential Chunk Loop

Process chunk00.txt, chunk01.txt, ... one at a time.

CRITICAL: Chunk numbers are ALWAYS zero-padded to two digits (chunk02.txt, NOT chunk2.txt).

For each chunk:

1. Read: [WORKSPACE]/chunks/chunk_XX.txt
2. Topic detection — keyword-scan chunk text for game/royal/tokusatsu signals:

``bash # Scan keywords directly (no dedicated script — detecttopics.py deprecated) grep -iE "FGO|Fate|Arknights|ไรเดอร์|Rider|เซนไต|Sentai|royal|imu|112" "[WORKSPACE]/chunks/chunkXX.txt" ` Use grep for speed; only install detect_signals.py` pipeline for bulk analysis.

1. DB check — ONLY if topic scan shows game keywords (FGO, Fate, YGO, 遊戯王):

- python3 "[SKILLROOT]/scripts/fetchfgodb.py" --check - python3 "[SKILLROOT]/scripts/fetchygodb.py" --check - Exit code 1 → re-run without --check to build DB. Exit code 0 → skip.

1. Generate timestamps: follow the Prompt Template below. DO NOT output the markdown into chat.
2. Write to [WORKSPACE]/chunkoutputs/chunkXX_output.md using the write tool.
3. Update State (CRITICAL): IMMEDIATELY overwrite [WORKSPACE]/anibontimestamperstate.json. Set "current_chunk" to XX+1. Do this after EVERY chunk.
4. End Turn (CRITICAL): Stop immediately after state update. Output [CHUNK COMPLETE. READY FOR NEXT.] and wait for the user to prompt you.
5. Handoff — if overwhelmed, write state file and stop:

{
  "video_id": "VIDEO_ID",
  "video_url": "VIDEO_URL",
  "workspace_path": "/absolute/path/to/youtube_VIDEO_ID_workspace",
  "total_chunks": 48,
  "current_chunk": 12,
  "db_checked": { "fgo": true, "ygo": false },
  "phase": "chunk_loop",
  "last_updated": "2026-07-17T09:23:00Z"
}

CRITICAL NOTE on currentchunk: Set it to the NEXT chunk to process, NOT the one you just finished. If you finished chunk11, write "current_chunk": 12.

📄 Prompt Template

Read the chunk. Group consecutive lines that discuss the same topic into one block. Write a short header title (in the output language confirmed in Step 1), then list timestamps below.

Rules:

• SUMMARIZE, DO NOT TRANSCRIBE. Combine multiple dialogue lines into one summary sentence.
• MAX 10 LINES per chunk. Outputting 30 lines = failure.
• Use HH:MM:SS directly from the file. Do NOT recalculate timestamps.
• Skip any line whose timestamp > the cutoff in the chunk header.
• Same topic within 1–2 minutes → one block, one header.
• New topic → new header. That is the only decision you need to make.
• Description: what was actually said/done (in chosen language). No internal feelings.
• If nothing notable: one line → HH:MM:SS ไม่มีเหตุการณ์สำคัญ

Output format for each chunk file:

<!-- chunk_00 | 00:00:00 – 00:05:00 -->

### ทักทาย
00:03:52 -  บ๊อตทักทายผู้ชม เริ่มสตรีม
00:04:01 -  พูดถึงช่วงเว้นว่างจากข่าวการเมือง

### หัวข้อถัดไป
HH:MM:SS -  description

• First line: HTML comment (required for assembly merge).
• ### Title on its own line before each topic block.
• HH:MM:SS - description — timestamp, dash, TWO spaces, text.
• No meta-commentary or apologies in the output file.

Edge cases:

• Trust item.start from the file — timestamps come from YouTube captions.
• Gap > 10 min with no transcript data → write HH:MM:SS [GAP: no transcript data]

Step 4: Assembly

Concatenate all chunk outputs:

Mac/Linux:

cat "[WORKSPACE]/chunk_outputs/chunk_"*"_output.md" > "[WORKSPACE]/raw_timestamps.txt"

Windows (PowerShell):

Get-Content (Get-Item "[WORKSPACE]/chunk_outputs/chunk_*_output.md" | Sort-Object Name) | Set-Content "[WORKSPACE]/raw_timestamps.txt"

Build [WORKSPACE]/parts.json from raw_timestamps.txt — one entry per section:

[
  {
    "title": "ทักทาย",
    "start": "00:03:52",
    "body": "00:03:52 -  บ๊อตทักทายผู้ชม เริ่มสตรีม\n00:04:01 -  พูดถึงช่วงเว้นว่างจากข่าวการเมือง"
  }
]

Then run the packer:

python3 "[SKILL_ROOT]/scripts/pack_timestamps.py" "[WORKSPACE]/timestamps.txt" --output "[WORKSPACE]/anibon_timestamps.md"

Output: [WORKSPACE]/anibontimestamps.md + [WORKSPACE]/timestampsparts.json

If any section exceeds 15 timestamp lines, adjust --byte-limit or split the timestamp list, then re-run.

Step 5: Verify

python3 "[SKILL_ROOT]/scripts/check_sections.py" "[WORKSPACE]/anibon_timestamps.md"

Any ❌ or ⚠️ → adjust --byte-limit or split timestamps → re-run pack_timestamps.py → re-verify. Do not proceed until all sections pass.

Iron Rules (Local Edition)

• ONE chunk per turn: No batch processing. Ever.
• Write tool, not chat: Never paste markdown into the conversation.
• State after every chunk: If you crash, state file is your recovery.
• No ls: You know the paths. Use them.
• No vision: Local models use --format txt, not --vision.
• Handoff over crash: If context > 10%, save state and hand off. Do not power through.
• Use grep for topic scan: detect_topics.py deprecated/deleted. Use grep -iE on chunk text instead.