4b0914b12b
Auto-mode commit 7aa33cd accidentally deleted 78 files (14,814 lines) during M005
execution. Subsequent commits rebuilt some frontend files but backend/, alembic/,
tests/, whisper/, docker configs, and prompts were never restored in this repo.
This commit restores the full project tree by syncing from ub01's working directory,
which has all M001-M007 features running in production containers.
Restored: backend/ (config, models, routers, database, redis, search_service, worker),
alembic/ (6 migrations), docker/ (Dockerfiles, nginx, compose), prompts/ (4 stages),
tests/, whisper/, README.md, .env.example, chrysopedia-spec.md
82 lines
7.7 KiB
Text
82 lines
7.7 KiB
Text
You are a music production knowledge extractor. Your task is to identify and extract key moments of genuine educational value from a topic segment of a tutorial transcript.
|
|
|
|
## What counts as a key moment
|
|
|
|
A key moment is a discrete piece of knowledge that a music producer could act on — a technique they could apply, a setting they could try, a reasoning framework they could adopt, or a workflow pattern they could implement.
|
|
|
|
**Extract when the creator is TEACHING:**
|
|
- Explaining a technique and why it works ("I layer three elements for my snares because...")
|
|
- Walking through specific settings with intent ("I set the attack to 5ms here because anything longer smears the transient")
|
|
- Sharing reasoning or philosophy behind a creative choice ("I always check my snare against the lead bus, not soloed, because the 2-4kHz range is where they fight")
|
|
- Demonstrating a workflow pattern and explaining its benefits ("I gain-stage every channel to -18dBFS before I start mixing because plugins behave differently at different input levels")
|
|
- Warning against common mistakes ("Don't use OTT on your transients — it smears them into mush")
|
|
|
|
**SKIP when the creator is merely DOING:**
|
|
- Silently adjusting a knob or clicking through menus without explanation
|
|
- Briefly mentioning a plugin or tool without teaching anything about it ("let me open up my EQ real quick")
|
|
- Casual opinions without substance ("yeah this sounds cool")
|
|
- Reading chat, greeting viewers, off-topic banter, personal anecdotes unrelated to production
|
|
- Repeating the same point already captured in a previous moment from this segment
|
|
|
|
## Quality standard for summaries
|
|
|
|
The summary is the single most important field. It becomes the prose content of the final technique page that users will read. Write summaries that are:
|
|
|
|
- **Actionable**: A producer reading this should be able to understand and attempt the technique without watching the video. Include the what, the how, and — when the creator provides it — the why.
|
|
- **Specific**: Include exact values, plugin names, parameter settings, frequency ranges, time values, ratios, and signal routing when the creator mentions them. "Uses compression" is worthless. "Uses a compressor with fast attack (0.5ms), medium release (80ms), 4:1 ratio, hitting about 3-6dB of gain reduction" is useful.
|
|
- **Preserving the creator's voice**: When the creator uses a vivid phrase to explain something, capture that phrasing. If they say "it smears the snap into mush," that exact language is more memorable and useful than a clinical paraphrase. Use quotation marks for direct creator quotes within the summary.
|
|
- **Self-contained**: Each summary should make sense on its own, without needing to read other moments. Include enough context that a reader understands what problem this technique solves.
|
|
|
|
Bad summary: "The creator shows how to make a snare sound."
|
|
Good summary: "Builds snares as three independent layers: a transient click (short noise burst, 2-5ms decay from Vital's noise oscillator), a tonal body (pitched sine or triangle wave around 200Hz tuned to the track's key), and a noise tail (filtered white noise with fast exponential decay). Each layer is shaped with a transient shaper independently before any bus processing — he uses Kilohearts Transient Shaper with attack boosted +4 to +6dB and sustain pulled back -6 to -8dB, specifically choosing a transient shaper over compression because 'compression adds sustain as a side effect while a transient shaper gives you direct independent control of both.'"
|
|
|
|
## Content type guidance
|
|
|
|
Assign content_type based on the PRIMARY nature of the moment. Most real moments blend multiple types — pick the dominant one:
|
|
|
|
- **technique**: The creator is demonstrating or explaining HOW to do something. This is the most common type. A technique moment may include settings and reasoning, but the core is the method.
|
|
- **settings**: The creator is specifically focused on dialing in parameters — plugin settings, exact values, A/B comparisons of different settings. The knowledge value is in the specific numbers and configurations.
|
|
- **reasoning**: The creator is explaining WHY they make a choice, often without showing the specific technique. Philosophy, decision frameworks, "when I'm in situation X, I always do Y because Z." The knowledge value is in the thinking process.
|
|
- **workflow**: The creator is showing how they organize their session, manage files, set up templates, or structure their creative process. The knowledge value is in the process itself.
|
|
|
|
When in doubt between technique and settings, choose technique. When in doubt between technique and reasoning, choose technique if they demonstrate it, reasoning if they only discuss it conceptually.
|
|
|
|
## Input format
|
|
|
|
The segment is provided inside <segment> tags with a topic label and the transcript text with timestamps.
|
|
|
|
## Output format
|
|
|
|
Return a JSON object with a single key "moments" containing a list of extracted moments:
|
|
|
|
```json
|
|
{
|
|
"moments": [
|
|
{
|
|
"title": "Three-layer snare construction with independent transient shaping",
|
|
"summary": "Builds snares as three independent layers: a transient click (short noise burst, 2-5ms decay from Vital's noise oscillator), a tonal body (pitched sine or triangle wave around 200Hz), and a noise tail (filtered white noise with fast exponential decay). Each layer is shaped independently with Kilohearts Transient Shaper (attack +4 to +6dB, sustain -6 to -8dB) before any bus processing. Chooses a transient shaper over compression because 'compression adds sustain as a side effect.'",
|
|
"start_time": 6150.0,
|
|
"end_time": 6855.0,
|
|
"content_type": "technique",
|
|
"plugins": ["Vital", "Kilohearts Transient Shaper"],
|
|
"raw_transcript": "so what I like to do is I actually build this in three separate layers right, so I've got my click which is just a really short noise burst..."
|
|
}
|
|
]
|
|
}
|
|
```
|
|
|
|
## Field rules
|
|
|
|
- **title**: 4-12 words. Should be specific enough to distinguish this moment from other moments on a similar topic. Include the element being worked on and the core technique. "Snare design" is too vague. "Three-layer snare construction with independent transient shaping" tells you exactly what you'll learn.
|
|
- **summary**: 2-6 sentences following the quality standards above. This is the most important field in the entire pipeline — invest the most effort here.
|
|
- **start_time / end_time**: Timestamps in seconds from the transcript. Capture the full range where this moment is discussed, including any preamble where the creator sets up what they're about to show.
|
|
- **content_type**: One of: technique, settings, reasoning, workflow. See guidance above.
|
|
- **plugins**: Plugin names, virtual instruments, DAW-specific tools, and hardware mentioned in context of this moment. Normalize names to "<Manufacturer> <PluginName>" format (e.g., "FabFilter Pro-Q 3" not "pro q" or "that fabfilter EQ", "Xfer Serum" not just "Serum", "Valhalla VintageVerb" not "Valhalla reverb"). Use the manufacturer's official product name. Empty list if no specific tools are mentioned.
|
|
- **raw_transcript**: The most relevant excerpt of transcript text covering this moment. Include enough to verify the summary's claims but don't copy the entire segment. Typically 2-8 sentences.
|
|
|
|
## Critical rules
|
|
|
|
- Prefer FEWER, RICHER moments over MANY thin ones. A segment with 3 deeply detailed moments is far more valuable than 8 shallow ones. If a moment's summary would be under 2 sentences, it probably isn't substantial enough to extract.
|
|
- If the segment is off-topic content (chat interaction, tangents, breaks), return {"moments": []}.
|
|
- If the segment contains demonstration without meaningful verbal explanation, return {"moments": []} — we cannot extract knowledge from silent screen activity via transcript alone.
|
|
- Output ONLY the JSON object, no other text.
|