AGENT SKILL

v2-vlog

Use for every edit whose project theme is "learning" (v2 Live Motion), before the cut spec, before the first cue and before calling the edit finished. The rules, the technique menu, the brand and the bar a learning-vlog edit must clear.

License 许可
Apache-2.0
最近更新
2026-11-06
Stars
★ 1
Forks
0

01 前言 Front Matter

name
v2-vlog
description
Use for every edit whose project theme is "learning" (v2 Live Motion), before the cut spec, before the first cue and before calling the edit finished. The rules, the technique menu, the brand and the bar a learning-vlog edit must clear.

02 资源文件 1

03 SKILL.md 正文

<!-- Created by The Angelo Umali. -->

v2 Live Motion learning vlog

The learning theme is the house standard for a talking learning vlog: the edit proves the script. Every thing the speaker names, every number he says and every claim he makes gets its visual on the word. Read this whole skill before set_cuts, then follow edit-director for the flow and cut-spec, cues, captions, reels and qc-gates for the field details.

Before the cut: ask about other people

When the footage shows anyone besides the creator, ask the user whether their faces are blurred. Never assume: some videos blur everyone else, others do not. Put the answer in the cut spec's blurFaces ("others" or "none"); a render of footage with more than one person refuses without it. Use a question card if you have one, otherwise ask in plain words and wait for the answer.

The rules (an edit that breaks one is not done)

  1. The edit proves the script. Before any cue, write the beat map with set_beat_map: one row per beat with the word, its source time, the visual and whether it sits on footage or paper. Every cue comes from a row.
  2. No backdrops. No plate behind a word or a number. Chips are labels, never backgrounds.
  3. No empty paper. Every second of a paperScene carries a graphic that explains the line (stepBars, listRows, a bigWord, a count). A heading alone on cream is a fault.
  4. Nothing on his face (eyes to chin, plus the forehead). Call frame_faces and keep every graphic above the face box minus about 80 px for the hairline, in the top band (baselines y 170–540 on the 1080 grid). Captions sit at y 1750, below his mouth and chin. qc_master fails the edit on any graphic pixel over the face.
  5. All text stays fully visible. No word or number goes behind his head, hair or face, and none covers his face. Resize it or move it into clear space, usually the top band, or leave it out. Text never overlaps text: a chip, a big word's kicker, a heading and a caption each need their own space at the same moment. A big word's kicker sits just above the word, so it shares the top band with any chipRail on screen with it: use one or the other. render_master refuses colliding text before it draws a frame, and qc_master checks it again (textClear).
  6. Never invent. A prompt, chat, receipt or call to action appears only when he says or did that thing, in his words. Counts start from 0 or from a value he says.
  7. No pictograms. No stick figures, outline icons, clip-art or invented charts and numbers. Every visual is his own footage, or a real UI rebuilt and animated. Money and trading screens are generic: no brand, balance, ticker or price.
  8. Map every place. A vlog that leaves home pins each place it visits on a real map (routeMap) the first time the video arrives there, with the name a viewer knows. Only proven places: the clip's own GPS, or a place the user names (the cut spec's places). Never home, so a video filmed at home has no map. qc_master fails an unpinned place (map).

Outputs

Unless the user asks for one output only, make two:

  • The full edit: the complete substantive story, with no length cap.
  • A reel of 120–180 seconds (set_reel, see reels), counted on the finished reel, branding and ending included. Open it on the strongest hook. Never pad or loop to reach the minimum, and never call a highlight cut the full edit.

Each output passes qc_master and is exported on its own target (full, reel).

Max creativity

  • Variety. No technique carries more than 30 % of the headline beats, and none runs three in a row. Chips, paper and captions ride along and do not count as beats. qc_master checks it.
  • A moving camera. The footage moves on the words that matter, as cut-spec zooms: lean where his face is already close, push for weight, snap for a punchline, through-out then through-in across a cut, reveal to open wide after a paper scene. No static stretch over about 6 s of footage.
  • Rack focus. Where he turns inward ("focus", "what I learned"), add a cut-spec focus window so the room blurs behind him.
  • Live motion over words. When he names a thing, show it working (the menu below). A word on screen is the fallback.
  • He stays on screen. For a paper explainer longer than 3 s, shrink his footage into a footageCard beside it (shrink in, grow out), alternating sides beat to beat.
  • Music by default. Set the cut spec's music sections (cold, intro, chapter, outro, with an energy each) across both outputs. The bed sits under his voice and ducks while he talks. Leave it out only when the user says so.
  • Honesty. If the edit is not at this bar, say so and say what is missing. Never call it finished when it is not.

Framing wide footage

A square or landscape source arrives letterboxed, and a letterboxed second shows the whole room. Set the cut spec's follow: "speaker": every kept second you have not framed yourself gets a 9:16 window on whoever is speaking, switching when the speaker changes, and holding the last speaker while the voice comes from off screen. Add your own reframes windows only where a shot needs another framing (cx, cy as shares of the picture); yours win. A source that already fills 9:16 needs neither. qc_master fails a letterboxed second that is still in the edit (framed).

The technique menu, by job

He says… Use
the one word that carries the line bigWord with drop or tumble in clear space, orange for the turn
two words that swap bigWord slot (a fixed line and the rolling words)
something he rules out bigWord scribbleAt or strikeAt
a number or an age bigWord count in clear space, with a kicker chip
a feeling or a level meterGauge, the needle settling on its word
a person or moment he mentions footageCard with sourceSecond and a crop that keeps the whole head
money, saving, spending phoneScreen (savings, allowance or cart), his card beside it
asking AI, a question and its answer chatThread: the question types and sends, the reply streams in his words
typing a prompt or a command typedPrompt or askField
markets, trading tradeScreen (live candles, no figures)
a tally receiptPrint, the total landing orange
a process flowSteps (check, coins, unlock)
a comparison that rises paperScene + stepBars (no numbers) + arrow
a list he counts off paperScene + listRows (card: true with a note), one pin per item
a place he goes, a trip routeMap: the real map, a dashed gold route drawing stop to stop on the word
a short label for the beat chipRail at y: 190, centred, alone in the top band
the open and the loop paperScene with logo, and a footageCard that grows or shrinks
the closing ask, only if he makes one commentCta

Call list_cue_kinds for each kind's fields and width caps. Every at and *At inside a cue is seconds after the cue's at, landing on its word.

The brand

Paper #F3EFE8, navy ink #172230, gold #C9A838 for fills (chips, pills, bars, checks, strikes, edges, arrows) with navy text on it, and orange #EA580C for a highlighted word. Gold text is too light; never steel blue. Chakra Petch sets display type and captions, JetBrains Mono the chips. The opening paperScene carries the logo and hosts the brand intro (intro: true, introAt on the first natural pause after the hook). Set the cut spec's grade to "learning".

Captions

The render draws word captions on a learning project. Call make_caption_phrases, then set_captions with each entry's keyword pills (words of its en) and a scale between 0.6 and 2.0 for the line that needs weight.

Done means

  1. qc_master passes every blocking gate on both outputs: frames, loudness, true peak, motion, density, variety, face clear, text clear, map, and on wide footage framed (and faces blurred when blurFaces is "others").
  2. A preview pass at every beat (render_preview, four stills a call) finds nothing left from the fault list in references/critique.md, on the full edit and the reel.
  3. Then export_deliverable for each output, and suggest_lessons for what the user corrected.

Hands to next: a beat map, a cut spec with the blur answer, follow, zooms, focus, grade and music, a learning cue sheet and a reel → cut-spec, cues, captions and reels, then qc-gates.

04 同仓库更多 Skills