--- name: anibon-timestamper-local description: Use when generating timestamps for Anibon streams on a LOCAL model (Ollama, gemma4, qwen). Focuses on step-by-step sequential looping, minimal context, and goldfish brain rules. --- # 🎬 Anibon Timestamper (Local AI Edition) Optimized for local LLMs with limited context windows (e.g. Gemma 4, Qwen 2.5). Processes video transcripts in sequential chunks to preserve memory. --- ## 🧠 Goldfish Brain Rules (CRITICAL) **Violating the letter of these rules is violating the spirit of these rules.** ### Red Flags - STOP and Start Over If you catch yourself doing any of the following, STOP GENERATING TEXT AND CALL A TOOL: - Generating the words "Wait", "Actually", or "Hold on" (indicates an 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 the generated timestamps/markdown directly into your response text instead of calling the `write` tool. ### Anti-Rationalization Table | Your Excuse | The Reality | |-------------|-------------| | "I'll do 18 and 19 in one turn" | You will crash the system. **Strictly ONE chunk per turn.** | | "I need to check the folder first" | Curiosity wastes context. Blindly execute the exact paths provided. | | "I am on Windows, I'll use `\`" | Backslashes break tool calls. **Always use forward slashes (`/`).** | | "I'll narrate my tool call" | No dry-running. If you say "I will read", you MUST actually call the tool. | | "I'll output everything to be safe" | The Prompt Template strictly forbids >10 lines. You MUST summarize. | | "I successfully generated the markdown" | Did you actually call the `write` tool? NEVER skip the `write` tool call! | | "I must continue reading the next chunk" | NO! You MUST stop after updating the state. Let the user prompt you. | | "I'll update the state file at the end of the batch" | No! If you batch 2 chunks, you MUST update the state file after EACH chunk. | ### Core Operating Constraints 1. **One tool per turn**: Never run batch actions. 2. **Process inline**: Read `.txt`, write timestamps yourself. No subagents. 3. **No `` tags**: Never wrap thoughts in ``. 4. **Handoff/Emergency Stop**: If user says "handoff", "stuck", or "% > 10", IMMEDIATELY write `anibon_timestamper_state.json` and halt. --- ## 🗺️ Plugin Directory Map (Do NOT use `ls`) You already know where everything is. Never search for files. - **Main Scripts**: `[PLUGIN_ROOT]/scripts/prepare_video.py` - **Sub-scripts**: `[PLUGIN_ROOT]/skills/anibon-timestamper/scripts/` (`fetch_fgo_db.py`, `fetch_ygo_db.py`, `check_sections.py`) - **Workspace**: `C:/Users/peter/youtube__workspace/` - `/chunks/chunk_XX.txt` - `/chunk_outputs/chunk_XX_output.md` - `/anibon_timestamper_state.json` --- ## 🧭 Steps > [!IMPORTANT] > **REQUIRED SUB-SKILL (FIRST):** Before anything else, run `preparing-tools` to verify all system dependencies (`yt-dlp`, `ffmpeg`, `python3`, `sqlite3`). Do NOT proceed if any tool is missing. ### Step 1: Resolve Plugin Path & Verify Shell is MINGW64/bash. Run: 1. **Find Plugin Root**: Look at the `` tag at the top of your prompt. Delete the exact text `skills\anibon-timestamper-local\SKILL.md` from the end of it. Replace all backslashes `\` with forward slashes `/`. This remaining path is your `[PLUGIN_ROOT]`. - 🚨 **ANTI-TYPO WARNING**: You (the AI) are prone to confusing hyphens (`-`) and underscores (`_`) when typing from memory. The repository is `anibon-stream-synthesis` (HYPHENS). The skill folder is `anibon-timestamper` (HYPHENS). NEVER use underscores for these! Copy the path directly, do not retype it. 2. **Verify**: ```bash uname -a && python3 --version ``` ### Step 2: Download & Chunk ```bash python3 "[PLUGIN_ROOT]/scripts/prepare_video.py" "VIDEO_URL" --format txt --block 300 --overlap 30 ``` *(If blocked by YouTube, ask user for cookies or raw_transcript.json.)* ### Step 3: Sequential Chunk Loop For each chunk `chunk_00.txt`, `chunk_01.txt`, ...: **CRITICAL**: Chunk numbers are ALWAYS zero-padded to two digits (e.g. `chunk_02.txt`, NOT `chunk_2.txt`). 1. **Read**: `read C:/Users/peter//chunks/chunk_XX.txt` 2. **DB check** (only if chunk text contains `FGO`, `Fate`, `YGO`, or `遊戯王`): - FGO: `python3 "[PLUGIN_ROOT]/skills/anibon-timestamper/scripts/fetch_fgo_db.py" --check` - YGO: `python3 "[PLUGIN_ROOT]/skills/anibon-timestamper/scripts/fetch_ygo_db.py" --check` - Exit 1 → re-run without `--check` to build. Otherwise skip. 3. **Generate timestamps**: follow the Prompt Template below. **DO NOT output the markdown directly into the chat!** 4. **Write**: Use the `write` tool to save the generated markdown directly to `C:/Users/peter//chunk_outputs/chunk_XX_output.md`. 5. **Update State (CRITICAL)**: IMMEDIATELY overwrite `anibon_timestamper_state.json` and set `"current_chunk"` to the NEXT chunk number (XX+1). Doing this after every single chunk ensures your progress is never lost if you crash. 6. **End Turn (CRITICAL)**: You MUST stop generating text immediately after updating the `.json` state. Do NOT call the `read` tool for the next chunk. Tell the user: `[CHUNK COMPLETE. READY FOR NEXT.]` and wait for their prompt. 7. **Handoff** if overwhelmed — write state file: ```json { "video_id": "VIDEO_ID", "video_url": "VIDEO_URL", "workspace_path": "C:/Users/peter/youtube__workspace", "total_chunks": 48, "current_chunk": 12, "db_checked": { "fgo": true, "ygo": false }, "phase": "chunk_loop", "last_updated": "2026-06-29T09:23:00Z" } ``` --- ### 📄 Prompt Template Read the chunk. Group consecutive lines that discuss the **same topic** into a block. Write a **short Thai header title** for each block, then list timestamps below it. **Rules:** - **CRITICAL: SUMMARIZE, DO NOT TRANSCRIBE.** Do not copy every line! You MUST combine multiple lines of dialogue into a single summary sentence. - **MAX 10 LINES**: You are strictly forbidden from writing more than 10 timestamp lines per chunk. If you output 30 lines, you have failed. - Use `(HH:MM:SS)` directly from the file. Do NOT recalculate. - 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's the only decision you need to make. - Description: what was actually said/done (Thai). No internal feelings. - If nothing notable: one line → `HH:MM:SS ไม่มีเหตุการณ์สำคัญ` **Output format for each chunk file:** ``` ### ทักทาย 00:03:52 - บ๊อตทักทายผู้ชม เริ่มสตรีม 00:04:01 - พูดถึงช่วงเว้นว่างจากข่าวการเมืองในช่วง 3 สัปดาห์ที่ผ่านมา ### หัวข้อถัดไป HH:MM:SS - description ``` - First line: HTML comment (required for merge). - `### Title` on its own line before each topic block. (This is a temporary header for grouping). - `HH:MM:SS - description` — timestamp, dash, two spaces, Thai text. No tags. - No meta-commentary or apologies in this file. **Edge cases (check only if gap > 10 min or timestamps seem wrong):** - Trust `item.start` over your intuition — timestamps come from YouTube captions. - Gap > 10 min with nothing in between → write `HH:MM:SS [GAP: no transcript data]`. --- ### Step 4: Assembly ```bash cat chunk_outputs/chunk_*_output.md > raw_timestamps.txt ``` Assemble `timestamp_VIDEO_ID.md`. Read through `raw_timestamps.txt` and convert the temporary `### Title` blocks into the final format. If a section exceeds **15 timestamp lines** (Thai characters consume 3x bytes and hit YouTube limits faster), split it: duplicate the header as Part 1 / Part 2 and divide the lines between them. For each section, use this exact format: ``` 📌 ส่วนที่ N: [1 sentence summary of what happened in this section] (หัวข้อ: [short title from the ### header] | ⏱ เริ่ม: HH:MM:SS) ------- [timestamps here — e.g. 00:03:52 - description] ``` - Use a running counter `N` across the whole file. - `HH:MM:SS` in the header must be the timestamp of the VERY FIRST event in this section. - Remove all `` markers and `###` headers. ### Step 5: Verify ```bash python3 "[PLUGIN_ROOT]/skills/anibon-timestamper/scripts/check_sections.py" timestamp_VIDEO_ID.md ``` Split further if any section fails (❌ or ⚠️).