@autoaiai/aicut 0.1.11 → 0.1.13

package/skills/lanshu-create-ai-presenter-video/references/editing.md

@@ -1,67 +0,0 @@
-# Editing
-
-Read this file for the visual plan, deterministic timeline, openings, closes, screen demos, captions, keyword callouts, covers, previews, and exports.
-
-## Visual routes
-
-- **Presenter-led:** the presenter carries the explanation; add one concise visual idea per beat.
-- **Screen-demo:** preserve real UI readability; introduce the presenter large at chapter starts, then shrink the same continuous video into a verified safe region.
-- **Mixed explainer:** assign supporting media only where it proves or clarifies narration.
-
-Do not fabricate product evidence. Label conceptual visuals clearly.
-
-## Timeline contract
-
-The locked narration defines duration and semantic boundaries. Every clip keeps three independent values:
-
-- authored start in the final video;
-- authored duration;
-- source-media start.
-
-When an opening changes length, shift authored starts while preserving presenter source offsets. Keep video sources muted and use approved external narration as the program clock. Prefer one continuous presenter source and source-time slices over regenerated chapter clips.
-
-When the presenter was generated in segments, place each one at its planned `authored_start_s` with source start `0` and authored duration equal to its audio slice. Independent segments leave small expression or pose differences at each seam. Mask every seam with a short opacity breath of about `0.2s`, a planned layout or shape change, or a cut to supporting media, and inspect frames at seam `±0.1s`.
-
-Use hard cuts at settled semantic pauses. Add transitions only when they clarify structure. Preserve a quiet tail so the last word, gesture, and mouth position finish naturally.
-
-## Opening, body, and close
-
-- Establish the topic and payoff within the first few seconds.
-- Design the time-zero frame as a deliberate cover frame.
-- Keep opening copy away from eyes, mouth, and the hand-motion path.
-- Use the same visual system across the cover, chapters, captions, keywords, progress, and close.
-- End with a conclusion or one requested next step. Do not add promotional copy without a brief requirement.
-
-## Captions
-
-Generate word timestamps from the final audio. Split into short semantic phrases, normally one readable line. A practical baseline is about `40ms` visual lead and `120ms` tail hold when neighboring phrases allow it.
-
-Keep captions clear of the face, UI, watermark, progress line, and platform controls. Highlight one meaningful term per phrase. Make entrances and exits quick enough to preserve reading stability.
-
-## Presenter-side keyword presets
-
-Use keyword callouts to amplify selected spoken beats. Avoid repeating one identical card throughout the video.
-
-Rotate a small family of presets such as:
-
-- radial burst;
-- tilted ribbon or sticker;
-- hand-drawn circle or marker stroke;
-- large/small type contrast;
-- staggered word-chip cluster;
-- double-layer outline lockup.
-
-Bind each callout to a real spoken stress or ASR word anchor. Separate the entry timing of the kicker, main word, secondary word, marker, rays, and outline by a few frames. Let the callout settle briefly, then collapse before the next idea. Keep it inside a consistent presenter-side safe region and verify the complete motion path does not cover the face or hands.
-
-## Composition and preview
-
-Use a deterministic compositor available in the environment. For HyperFrames:
-
-1. load the current HyperFrames workflow and domain instructions;
-2. give all timed media explicit start, duration, track, and source offset;
-3. run the required composition check with representative emphasis and transition samples;
-4. inspect time zero, chapter cuts, layout changes, keyword beats, captions, close, and final frame;
-5. open the final Studio preview and obtain approval before rendering;
-6. render at delivery quality only after approval.
-
-Measure the assembled stereo program after rendering. Mono narration duplicated into stereo can measure about 3 LU louder, so perform final program normalization from the rendered file. `finalize_delivery.sh` does this and verifies the delivered master and share loudness before publishing either file.

package/skills/lanshu-create-ai-presenter-video/references/generation.md

@@ -1,94 +0,0 @@
-# Generation
-
-Read this file for intake, content, narration, capability selection, paid generation, presenter prompts, or provider changes.
-
-## Inputs and authorization
-
-Required:
-
-- a topic or script;
-- one local presenter image with one clear adult face;
-- confirmed image rights before remote upload;
-- confirmed adult status from the user.
-
-Optional:
-
-- authorized voice sample;
-- screen recordings, B-roll, images, charts, brand assets, or source links;
-- audience, duration, language, aspect, style, watermark, music, CTA, provider preference, or previous episode.
-
-Inspect the real image. Record visible anchors only: framing, hair, glasses, clothing, accessories, hands, furniture, background geometry, camera height, crop, light, exposure, white balance, embedded text, logos, occlusion, multiple faces, and resolution. Do not infer sensitive traits.
-
-When a voice sample exists, listen for one speaker, audible speech duration, language, noise, echo, clipping, music, silence, and codec damage. Record whether cloning is authorized. Keep the original unchanged.
-
-## Content and narration
-
-For a topic, use one clear spine:
-
-```text
-hook → promise → 2–4 useful beats → synthesis → close
-```
-
-For a supplied script, preserve factual meaning while improving breath, pronunciation, sentence length, and transitions. Flag material factual edits for approval.
-
-Generate the complete approved narration before visual work. Keep one voice identity and one synthesis configuration. Split only for operational limits or clean semantic sections. Preserve raw and normalized copies.
-
-Run ASR against final audio. Review names, numbers, English tokens, omitted words, repeated words, and tail speech. Use actual audio durations for every later cut.
-
-## Capability routing
-
-Resolve capabilities at execution time:
-
-| Capability | Acceptance priority |
-|---|---|
-| Voice generation | authorization, identity, pronunciation, rate control, clean audio |
-| Main presenter | exact audio support, semantic lip sync, identity stability, required duration |
-| Short motion | stable face, one controllable gesture, fixed camera, economical retry |
-| Lip-sync repair | preserves accepted motion while replacing mouth timing with exact audio |
-| Word-timestamp ASR | accurate word start/end times and reviewable output |
-| Timeline compositor | explicit source offsets, frame-accurate seek, deterministic render |
-| Encoder and QA | local probe, loudness, decode, black-frame and contact-sheet support |
-
-Inspect current installed tools and official documentation when availability, duration, pricing, or request fields may have changed. Prefer a user-selected provider when it passes the gates. Otherwise choose the simplest available capability that does.
-
-Record the actual provider, model, version, region, parameters, price evidence date, task ID, and sanitized request body in `job.json`. Ask before a provider change that affects cost, privacy, voice, appearance, or quality.
-
-## Billing and pilot gate
-
-Before the first remote or paid call, state:
-
-- files or data being uploaded;
-- capability and selected tool;
-- requested seconds or units;
-- known price or that the cost is unknown;
-- pilot duration and retry ceiling;
-- expected output and main risks.
-
-Generate the smallest useful pilot. Continue to a full run only after technical checks and visual review pass. Do not generate a long clip when the edit uses only a short range.
-
-## Duration caps and segments
-
-Prefer one continuous audio-driven presenter source. When the provider caps request duration or reference-audio length below the narration length, split by the narration's real pauses:
-
-- Use the smallest of the provider's video-duration and reference-audio limits as the cap.
-- Normalize the final-audio ASR sentence or word timings to `[{"start": s, "end": s, "text": "..."}]` and run `scripts/plan_segments.py`. It cuts only inside pauses of at least `0.25s`, takes the latest usable pause so the edit has as few seams as possible, and keeps every segment at least `2s` long.
-- With `--whole-seconds`, it prefers an integer cut that lies inside a pause; otherwise it cuts at the pause midpoint and requests the rounded-up seconds. Do not move a cut onto an integer that falls inside speech to save a billed second.
-- Send each segment exactly its planned audio slice. Reuse the same image, seed, framing, wardrobe, light, and motion constraints for the pilot and every segment.
-- Quote `total_requested_seconds` from the plan in the billing statement. Record the plan in `plan.segment_plan` and task IDs keyed by range, such as `"00-10": "<task id>"`.
-- Generated clips often run a few frames longer than requested. Trim them through authored duration and source offset; never time-stretch a presenter clip.
-
-## Presenter prompt structure
-
-Write prompts in this order:
-
-1. Bind the user image as the sole presenter and scene reference.
-2. List only visible identity, clothing, accessory, camera, framing, background, light, exposure, and white-balance anchors.
-3. Request realistic skin, hair, eye moisture, fabric, breathing, bilateral blinking, and restrained head motion.
-4. Supply exact dialogue through the tool's supported audio or dialogue field.
-5. Tie one physically plausible gesture to one phrase and approximate time.
-6. Reserve a settled, closed-mouth tail.
-7. Exclude identity, face, glasses, hair, wardrobe, skin tone, background, camera, lighting, finger, hand, dialogue, text, logo, watermark, and extra-person drift.
-
-For the main presenter, prioritize mouth timing and identity. Keep hands low or outside frame. For an opening or close, one small gesture may carry the hook or CTA. Keep hands below the collarbone, away from the face and lens, and end the movement before the tail hold.
-
-If the body performance is accepted but mouth timing is visibly late, preserve the motion plate and apply lip-sync repair with the exact locked audio and no duration extension.

package/skills/lanshu-create-ai-presenter-video/references/qa-recovery.md

@@ -1,83 +0,0 @@
-# QA and recovery
-
-Read this file before accepting generated media or delivery, and whenever a production fault appears.
-
-## Acceptance gates
-
-### Inputs
-
-- Topic or script exists.
-- Presenter image decodes and contains one reviewed adult face.
-- Image rights, adult confirmation, and remote-upload permission are recorded.
-- Any voice sample contains one authorized speaker.
-- Supporting media is probed and its provenance is recorded.
-
-### Audio
-
-- Final ASR matches the intended script once, without material omissions or additions.
-- Sections have consistent loudness, commonly near `-17 LUFS`.
-- Assembled program loudness is commonly `-16 ± 0.5 LUFS` unless the destination specifies another target.
-- No clipped words, doubled tracks, echo, clicks, unexpected silence, or truncated tail.
-- Measure the final stereo file; section measurements alone are insufficient. The finalizer records `master_loudness` and `share_loudness` and refuses to publish outside `±0.5 LU` or above `-1.0 dBTP`.
-
-### Presenter
-
-- Expected resolution, frame rate, duration, and decodable streams.
-- No black frames, long exact-frame freezes, progressive darkening, or duration extension.
-- Identity, face geometry, hair, glasses, clothing, accessories, background, crop, exposure, and white balance remain coherent.
-- Mouth follows names, numbers, English tokens, plosives, and phrase endings.
-- Blinks are sparse and bilateral; gestures occur once and settle; hands remain plausible and away from the face.
-- Tail ends with a settled face and resting mouth.
-- Segmented output keeps identity, framing, and light continuous at every seam `±0.1s`.
-
-### Composition and delivery
-
-- No flash, duplicate presenter, source-time reset, missing overlay, overflow, face obstruction, caption collision, or unreadable UI.
-- Opening, callouts, captions, watermark, progress, and platform safe zones remain compatible.
-- Master and share files have the expected duration, dimensions, frame rate, codec, pixel format, and audio rate.
-- Both files fully decode.
-- Every `freeze_events` entry in the delivery report corresponds to an intentional still shot.
-- Contact sheet covers opening, chapters, emphasis graphics, close, and final frame.
-- Watch the complete video at normal speed before delivery.
-
-## Recovery rules
-
-### Interrupted remote task
-
-Poll the saved task ID and download a succeeded result. Resubmit only after a confirmed failure or cancellation.
-
-### Voice identity or loudness changes
-
-Confirm one synthesis identity and configuration. Regenerate the mismatching section, normalize again, and rerun ASR. If the final render is louder than the source, confirm video is muted and apply whole-program two-pass loudness correction.
-
-### Presenter changes between chapters
-
-Use one continuous audio-driven source with source-time slices. When splitting is unavoidable, lock identity, framing, parameters, light, and quiet edge holds.
-
-### Body motion works but mouth is late
-
-Keep the accepted motion plate. Apply lip-sync repair with the exact locked audio and no duration extension. Check speech anchors at numbers, plosives, English tokens, and final syllables.
-
-### Visible jump at a segment seam
-
-Confirm each segment used its planned audio slice and identical generation settings. Mask small expression or pose differences with a short opacity breath, a layout change, or supporting media. Regenerate only the affected segment when identity, framing, or light differs structurally.
-
-### Finalizer rejects loudness
-
-Nothing was published. Confirm every video source is muted and narration is not routed twice, then re-render. When the destination requires another target, rerun with `PROGRAM_LUFS` set to it.
-
-### Face, glasses, hands, or light drift
-
-Retry from the original image with reduced motion and explicit camera, exposure, and white-balance locks. Simplify the gesture and keep hands low. Structural face or finger faults require regeneration; grading cannot correct them.
-
-### Frozen picture-in-picture
-
-Confirm the layout uses moving video, source offsets advance continuously, and no poster frame or identical-frame padding replaced the presenter.
-
-### Captions or keyword graphics feel wrong
-
-Regenerate timings from the final audio. Split by meaning, shorten visible phrases, and adjust small lead/hold margins. Bind keyword motion to actual spoken anchors and verify its complete path against the face, hands, UI, and captions.
-
-### Three rejected paid candidates
-
-Stop. Preserve candidates, prompts, task IDs, cost, and rejection notes. Summarize the recurring failure, remaining options, expected additional cost, and the single variable proposed for the next attempt.

package/skills/lanshu-create-ai-presenter-video/references/styled-explainer.md

@@ -1,189 +0,0 @@
-# Styled explainer route
-
-Read this file when the job's `creative.route` is `styled`: a performed explainer in one of twelve styles (nine visual
-styles and three narrative styles — 推演, 案卷, 小剧场 — whose look comes with a way of telling), with no presenter. The styles carry the explanation on their own — the work is in performing every line well. The engine
-lives in `explainer/`:
-`explainer/STYLES.md` to choose a style, `explainer/KITS.md` for the contract and tools, each style's
-`kits/<style>/starter/STARTER.md` for how to perform a beat, and `examples/kv-cache/<style>/` for a finished film.
-
-The input is a topic or script; the only paid generation is the voice. Films render 16:9 at 1920×1080 and end with a
-one-frame recap. For 9:16 vertical delivery use the presenter route. To bring a digital human into a styled film, see
-the last section — an optional method, not a separate route.
-
-## Requirements
-
-- Node.js with `npx` (HyperFrames is run as `npx --yes hyperframes@0.8.81`, override with `HYPERFRAMES`), FFmpeg,
-  `rsync`, Python 3.9+ with numpy for the sound bed (`PYTHON=…` to choose the interpreter), network access for
-  Google Fonts and jsDelivr (font subsets).
-- Voice and word timing: the built-in path calls MiniMax T2A once per line (`MINIMAX_API_KEY`; `MINIMAX_BASE_URL` for the
-  China endpoint) and takes each word's timestamp from the voice engine itself — no speech-recognition model to install.
-  Narration the user voiced or recorded elsewhere works too (`story.py --audio`): lines are found at its pauses and word
-  times are estimated (about 0.15 s), fine for captions and slightly looser for word-timed beats.
-
-### Voice setup · 配音
-
-The default voice is MiniMax `Chinese (Mandarin)_Reliable_Executive` (a steady male voice) at speed 1.1; set `voice:`
-and `speed:` in `script.md` to change it — the user picks a system voice ID from the voice library in their MiniMax
-console. The international API is `https://api.minimax.io` (the default); mainland-China accounts use their console's
-endpoint via `MINIMAX_BASE_URL` (typically `https://api.minimaxi.com`). Write names, numbers and acronyms in the
-script as they should be read; after voicing, check how numbers, units and English were read: `story/audio/pronounce.txt` lists each line as the engine read it
-(e.g. 「2 GB」 read as 「二吉字节」 — rewrite the script as 「两 GB」 if that is wanted, and voice that line again).
-Narration the user recorded themselves needs no TTS at all: `story.py story --audio`.
-
-## Job layout
-
-The usual job directory gains two folders:
-
-```text
-story/      script.md → audio/narration.wav, audio/tts/ (per-line voice + word timestamps), audio/pronounce.txt, story.json
-film/       the HyperFrames project (from the style's starter); film/beats.js holds the performed scenes
-```
-
-`X` below is `$SKILL_DIR/explainer`.
-
-## Production, state by state
-
-### intake
-
-```bash
-python3 "$SKILL_DIR/scripts/init_job.py" --job-dir ~/Videos/kv-cache --route styled \
-  --explainer-style v3-notebook --topic "讲清楚 KV cache"
-```
-
-If the user has not chosen a style, show them `explainer/STYLES.md` (attach or link its two gallery images) and
-recommend two or three from the topic and audience, one reason each; or build drafts in the nine visual styles (below)
-and let them pick from a still grid. Settle the style before the script is locked — the script is written for it (next
-step); for a narrative style (V10–V12), adapt its suggested arc to the topic.
-
-### content_locked
-
-Write `story/script.md` for the chosen style — tone, sentence length, `speed:` and the recap `length:` from the
-"按风格写稿" table in `explainer/STYLES.md` — with the content rules of [generation.md](generation.md): chapters (`## key | 章节标题`), one narration line per text line, `> beat notes` under each chapter
-(the starters show them in the draft), `## end` with `emphasis:` for the closing line, `## recap` points, and a
-`## facts` block — the only facts the film may put on screen. See `explainer/stories/kv-cache/script.md`.
-
-For a narrative style, these arcs are useful starting points; adapt chapter count and order to the explanation:
-V10 推演 a question → the figure built step by step → the pattern
-→ the result and its cost; V11 案卷 the anomaly and suspects → evidence → the cause proven → the fix → the case closed
-(cause / fix / cost / lesson, facts sourced, illustrative numbers marked 「示意」); V12 小剧场 one life scene and one
-character: the misconception → a bill, a queue or an object comparison that shows the cause → the new way → the analogy
-mapped to the real term, closed in one sentence. The bundled `explainer/stories/kv-cache-reasoning`, `kv-cache-casefile`
-and `kv-cache-sitcom` are the same topic written in the three arcs. A sitcom's setting and characters can change
-with the topic; the demo's milk-tea shop is an example. Keep facts sourced and label illustrative numbers.
-
-```bash
-python3 "$X/tools/story.py" story --check      # chapters, lines and the estimated length
-python3 "$X/tools/docs.py" story docs          # docs/SCRIPT.md and docs/BEAT_SHEET.md from script.md
-```
-
-Record `artifacts.script = docs/SCRIPT.md` and `artifacts.beat_sheet = docs/BEAT_SHEET.md`. Show the user the script and
-the estimated length before anything is voiced; that is their approval of the content and of the voice cost.
-
-### audio_locked
-
-`python3 "$X/tools/story.py" story` voices every line with MiniMax (cached per line, levelled to −18 LUFS so delivery
-loudness passes), or voice the script with any TTS — or use the user's own recording — and run
-`python3 "$X/tools/story.py" story --audio narration.wav` (the script text must match what is said).
-
-Read the printed word times against the audio; fix the script or voice before going on. Record
-`artifacts.final_audio = story/audio/narration.wav`, `artifacts.story = story/story.json`, and
-`qa.asr_report = story/story.json` (its per-word times are the ASR evidence).
-Read `story/audio/pronounce.txt` for misread numbers or terms. A `--dry` draft never locks the audio
-(`check_state.py` refuses it); `story.json` records `timing: engine | estimated | draft`.
-
-### visual_plan_locked
-
-Build the draft film — it plays end to end with captions, chapter chip, closing line, recap and sound, and a
-「待演绎」 placeholder per line:
-
-```bash
-bash "$X/tools/new_film.sh" <style> story film
-# to compare styles first:  DEST_ROOT=drafts JOBS=3 bash "$X/tools/new_topic.sh" story
-#                           python3 "$X/tools/still_grid.py" drafts 9.5,30 qa/contacts
-```
-
-Run `python3 "$X/tools/docs.py" story docs` again: it adds `docs/TIMELINE.md` and a `docs/STORYBOARD.md` table with one
-row per line (time and words). Fill its "on screen" column: what is performed, on which word, using the style's idioms
-(`film/STARTER.md`) and the finished example. Make t = 0 a composed frame (the first scene already on screen, not an
-empty stage) — it is the first thing a feed shows. Send the user the storyboard and a phone-friendly draft
-(`DRAFT=1 bash "$X/tools/render.sh" film <name> qa`, about a minute) rather than a developer preview; after their
-approval of style and storyboard, set `plan.status` to `"approved"`.
-
-### presenter_generated (here: the performed film)
-
-There is no presenter on this route; the state means the performed film exists. Replace the placeholders in
-`film/beats.js` with the storyboard's scenes, scene by scene, every beat time from `story.at()` / `story.phrase()` /
-`story.span()`. Rerun `python3 "$X/tools/fonts.py" film` and `(cd film && bash tools/sfx.sh)` after edits. Record
-`artifacts.film_project = film`, and write what you looked at and judged to a file (e.g. `qa/reports/visual-review.md`)
-and record its path as `qa.manual_visual_review` — the checker needs a file, not a sentence.
-
-### composition_checked
-
-```bash
-python3 "$X/tools/qa.py" film qa/film      # check + snapshots + stillness + cover.png → qa/film/composition.json
-```
-
-It exits non-zero when `check` fails or a stillness pair is under 0.001 (a parked camera or frozen picture). Look at
-the snapshots in `qa/film/contacts/` (every chapter, the closing line, the recap). `composition_file_too_large` and
-`nested_structure_needs_subcomposition` warnings are expected for these single-file films. Record
-`qa.composition_report = qa/film/composition.json`. Get the user's approval before the final render, with a new
-`DRAFT=1` render (or the Studio preview, if they prefer it).
-
-### rendered → verified
-
-```bash
-bash "$X/tools/render.sh" film <name> outputs        # renders film/renders/<name>.mp4, then finalize_delivery.sh
-```
-
-`render.sh` calls this skill's `finalize_delivery.sh` (loudness, full decode, contact sheet), writes
-`outputs/<name>.srt` (subtitles for Bilibili CC and other platforms) and `outputs/<name>-freeze.txt`; a performed
-explainer should have 0 freeze events. `SKIP_RENDER=1` repeats only the finalize step. Record
-`artifacts.rendered = film/renders/<name>.mp4`, `artifacts.master`, `artifacts.share`, `artifacts.caption_json` (or the
-SRT), `qa.delivery_report`, then `check_state.py --write` until `verified`. Hand the user the master, the share copy,
-the SRT, and `qa/film/cover.png` as the upload cover.
-
-Recording artifacts is plain JSON editing, e.g.
-`jq '.artifacts.film_project = "film"' job.json > job.tmp && mv job.tmp job.json`.
-
-## Re-styling an earlier film
-
-Keep the job's `story/` (script and locked audio — no new voice cost). Build a second film from the same story in the
-new style (`new_film.sh <style> story film-<style>`), perform its `beats.js` from the storyboard, and get the
-storyboard re-approved; QA and delivery as above, with its own output name. The first film stays untouched.
-
-## Performing the content
-
-- Act out every narrated phrase on screen, on its word: the thing named appears, moves, changes or is compared
-  when it is said. A page that only reveals text reads as a slide.
-- One plain-language note per chapter; the real term next to every analogy.
-- Only facts from the script and the `## facts` block; label illustrative visuals as illustrative.
-- Keep the style's aesthetic baseline (`explainer/KITS.md` §2): readable sizes at 1080p, one accent with one meaning,
-  no dead space, something meaningful in motion every second, the camera never parks.
-- Starters are drafts: never deliver a film whose `beats.js` still shows 「待演绎」 placeholders.
-
-## Optional: bringing a digital human into a styled film
-
-The styles are designed without a presenter, and most topics do not need one. When a user wants their digital human
-in a styled film anyway (a personal channel, a familiar face for the hook), this method was tried on two finished
-films (V2 and V9, both passing `check`); use it as a recipe, not a default.
-
-1. **One clock.** Lock the narration the presenter route's way and generate the presenter from it (pilot, cost
-   statement, evidence — [generation.md](generation.md)). Build the story from that exact audio:
-   `story.py story --audio assets/audio/final/narration.wav`. The film is then timed to the words the presenter's
-   mouth is already synced to; never re-voice one side only. Join a segmented presenter into one track first.
-2. **Give the presenter its own space — never cover the film.** Each style fills the frame with performed content;
-   anything laid on top hides it (and fails the layout and contrast checks). Shrink the film instead: wrap every
-   visual child of the composition root in CSS `translate` / `scale` (they compose with the style's own transforms)
-   and place the presenter in the freed column.
-3. **Three states, morphing over ~0.6 s:** *on stage* for the hook and the closing line (film at about 70 % on one
-   side, the presenter on a tall card beside it); *badge* while the film explains (film at about 86 %, the presenter as
-   a small circle in the narrow column); *off* for the recap (film full size). Default keys from the story: on stage
-   until the first line ends, badge until just before the closing line, on stage for it, off before the recap.
-4. **Belong to the style.** Frame the presenter with the style's tokens (`--kit-accent` border, `--kit-surface`
-   behind), the face focused near the top third (`object-position: 50% 30%`), the video muted (the film's narration
-   track is the presenter's audio).
-5. **Drive it from the film's clock.** Every frame is a pure function of t: update the presenter frame and the film
-   transform inside the style's frame function (or a hook it calls each frame), and give the `<video>` its own
-   `data-start` / `data-duration` so HyperFrames owns its playback.
-6. **Verify** as for any styled film (check 0 errors, stillness pairs), plus normal-speed playback of the hook and the
-   closing line for lip-sync.

package/skills/laoxu-video-script/SKILL.md

@@ -1,117 +0,0 @@
----
-name: laoxu-video-script
-description: "自媒体起号方法论（老徐教学版）——把「老徐」验证过的内容打法提炼成可迁移的方法，帮零基础学员定位账号、找选题、写出站得住的口播稿/内容、并自检迭代。当用户（尤其是学员/小白）要起号、做自媒体、找账号定位、想选题、写视频稿/口播稿/逐字稿、或要诊断改稿（钩子/逻辑/AI味/传播力）时使用。核心是「按观众问题顺序走」的三种骨架（问答讲解 / 推演解题 / 破立合评论）+ 实测题材优先 + 40 秒开场七式 + 母题库 + 四感及格线（新鲜感/有用感/有希望感/参与感）+ 写稿三忌 + 五维诊断 + 串联审稿。老徐本人的声音只是范例，学员用这套方法建**自己的**号。 Runs in AICUT: the film is film.html, a page per shot, drawn by `aicut look` and `aicut render`."
-license: MIT
-source: "chenmisss/laoxu-video-script"
----
-
-# 自媒体起号方法论 · 老徐教学版
-
-## Running it in AICUT
-
-AICUT's film is `film.html`: it cuts pages, footage, pictures and sound together on tracks. A page is a web page that
-sets `window.film = { frame(t) }` and draws that moment — no framework, no build step, no composition object, no
-`registerRoot`. Five shots are five pages cut together, not one page that changes scene, so each shot stays trimmable
-and replaceable in Studio.
-
-```sh
-npx -y @autoaiai/aicut@latest          # the manual: the format and the commands, always current
-npx -y @autoaiai/aicut@latest open .   # make this folder a film and print Studio's address
-npx -y @autoaiai/aicut@latest look     # a contact sheet of the whole film, its sound, and a determinism check
-npx -y @autoaiai/aicut@latest look 4.5 # one frame, full size — look at the PNG, not at the code
-npx -y @autoaiai/aicut@latest render   # the MP4, beside film.html
-```
-
-Run the manual before the first film: it is authoritative for the format, and this skill is about the method above it.
-
-Where the words below name another program's part, it becomes this:
-
-| elsewhere | in AICUT |
-|---|---|
-| composition, timeline, project file | `film.html`; tracks and clips inside it |
-| a scene, a sequence, a slide | a page (`pages/<name>.html`) cut onto a track |
-| `useCurrentFrame()`, `interpolate()` | `frame(t)` — the picture depends only on `t`, never on state or `Date.now()` |
-| a still, a screengrab | `aicut look <t>` |
-| `render` / `build` / `export` | `aicut render` (the person can also export from Studio) |
-| a dev server, a preview window | `aicut open` and the Studio address it prints |
-
-Hard rules that hold whatever the method says: the picture depends only on `t`; every value in the person's prompt is
-binding; look at the frames `look` writes before saying a film is done.
-
-
-把老徐做自媒体验证过的一套打法，提炼成**学员能照着用的方法**。目标不是让学员模仿"老徐"这个人，是让他们学会**这套方法**，去做出**自己的**号。
-
-## 这个 skill 教什么（也划清不教什么）
-
-**教**：账号定位、选题体系、把话题写成站得住的内容（破立合）、单篇成稿格式、以及像老徐定稿那样的**五维诊断 + 串联审稿**。适合零基础起号、也适合有号但内容没起色的人做体检。
-
-**老徐只是范例**：文档里凡出现老徐的人设/装备/签名，都是"一个定义清晰的账号长什么样"的**样板**，不是要你照抄。你的任务是套方法、填**你自己**的定位。
-
-**不教/不做**：平台注册流程、买量投流、封面（走 guizang-cover）、图表（走 转转图表 那套）、配图（走 xiaoxu-illustrations）、公众号排版（走 wechat-article）。这个 skill 只管**定位 + 选题 + 内容 + 诊断**这条主线。
-
-## 起号路线（学员照这个顺序走，一步都别跳）
-
-### 第 0 课 · 底层心法：为什么有人关注你，并留下来（先讲这个）
-先想明白"人凭什么关注你、凭什么留下来"再练招式：**勾 → 留 → 关注**三个动作、**让人占便宜**（你到底能分享什么）、**替身感 / 成长感**（为什么粉你这个人）、**活人感**（真实感受是护城河）、以及**起号前三篇**怎么打。这是"为什么"，后面五课是"怎么做"。见 `references/qihao-xinfa.md`。
-
-### 第 1 课 · 定位你自己的账号
-先想清「你是谁 × 给谁看 × 凭什么信你」。这一步没做完，后面全是无根之木。方法 + 老徐范例 + 学员自填工作表见 `references/persona.md`。
-
-### 第 2 课 · 建选题体系 + 你自己的母题库
-选题第一优先级是**实测题材**——「我拿真金白银替咱试了」（自然流量前三全是它；纯推演要靠逐点真案例撑，见文末标杆）。再学会把"人人都在吵的表层争论"往下挖成"底下那道真正的题"。选题/钩子/金句公式见 `references/formulas.md`；老徐的 **112 条母题原子库**（范例）见 `references/motifs-and-lines.md`，配套真稿例证 `references/cases.md`、机读版 `assets/atoms.jsonl`——学员照这套方法挖**自己领域**的母题库。
-
-### 第 3 课 · 写一篇站得住的稿
-三种骨架按题材选——**讲解 / 干货类**走「预判观众问题 → 回答 → 给解法」，**推演 / 解题类**立一道数字题当面解完（结尾拼成对号入座表），**评论类**走「破 → 立 → 用 → 推 → 合」（此形旗舰样例未发布无验证，前两种才是账号数据背书的主力）；底层都是"按观众的问题顺序往前走"。开场 40 秒内抛钩子，第一句起法看题材七选一（有自然播放背书的优先，反流量式是特定题材专用别当默认）。成稿过一条内容及格线——**四感**（新鲜感 / 有用感 / 有希望感 / 参与感）：前三感长在骨架上逐拍交付——**破立给新鲜、立用给有用、推合给希望**，参与感串在缝上（过桥句 / 照镜子句 / 收尾话头）；守三条语言铁律——**忌车轱辘话（句式复用也算）、忌过度形容词、忌报幕与弯弯绕**（动笔就守，别指望写完再洗）；创作视角**跟观众坐一边——多用「咱」、少用「你」**（「你」省给照镜子句那一两处）。方法读 `references/method.md`，成稿格式 + 空白模板见 `references/format.md` 和 `assets/template.md`。
-
-### 第 4 课 · 五维诊断（这是老徐和野路子的分水岭）
-写完不是完事，是**开始体检**。五个可单独调用的诊断——选题 / 开场钩子 / 逐字稿逻辑 / **AI 味扫描** / 传播力——见 `references/diagnostics.md`。学员也可以只拿一段来诊断（"帮我看看这个开头行不行"）。**去味改写（内置）**：D4 检测出 AI 味后，用内置模块 `references/de-ai.md` **去掉车轱辘话 / 空形容词 / AI 词汇**——它**改掉了通用去味工具里"删金句"这条会毁掉老徐风格的规则**（保留真金句、只杀假排比和啰嗦）。**已内置进本 skill**，不用再单独装。
-
-### 第 5 课 · 串联审稿 + 迭代留痕
-通查车轱辘话（含句式复用）、空形容词、报幕腔、称呼（你→咱）、过桥、首尾呼应、伏笔回收，发布前过**四感终检**（别处听过吗 / 明天能用吗 / 看完有劲吗 / 有话可接吗），改动写进「逐版修改记录」。见 `references/diagnostics.md` 末节。
-
-### 第 6 课 · 起号运营（把一篇变成一个号）
-选平台、定内容节奏、冷启动、留量漏斗、变现雏形——把方法接到"经营一个号"上。见 `references/qihao-path.md`。
-
-### 红线（教学时先讲）
-所有稿都守 `references/redlines.md`：批判结构不批判个人、对称分析、不消费旧关系、数字口径要标注。带学员时这条尤其重要——**帮他们躲开会翻车的坑**。
-
-## 参考文件导览（按需读，别一次全塞）
-
-| 文件 | 教哪一课 |
-|---|---|
-| `references/qihao-xinfa.md` | **底层心法**：勾/留/关注、让人占便宜、替身感/成长感/活人感、起号前三篇 |
-| `references/persona.md` | 定位：人设三问 + 老徐范例 + 学员工作表 |
-| `references/formulas.md` | 选题 / 开场钩子 / 金句 / 广告桥法 公式库 |
-| `references/motifs-and-lines.md` | **母题原子库 112 条**（老徐范例）+ 教你建自己的 |
-| `references/cases.md` | 案例库：20 条最强母题的真稿原文例证 |
-| `references/method.md` | 破立合骨架、**四感及格线**（新鲜/有用/希望/参与）、**写稿三忌**（车轱辘话/形容词/报幕弯弯绕）、**创作视角**（多咱少你）、过桥、扣环、情感曲线 |
-| `references/format.md` | 成稿格式的每一块 + 可粘贴模板 |
-| `references/diagnostics.md` | 五维诊断 + 串联审稿 |
-| `references/de-ai.md` | **去 AI 味模块**（去车轱辘话/空形容词/讲台腔，内置·带"保留金句"override） |
-| `references/redlines.md` | 立场与合规红线 |
-| `references/qihao-path.md` | 起号运营：平台分工 / 冷启动 / 留量漏斗 / 变现 |
-| `assets/template.md` | 空白成稿模板 |
-
-## 关于这个 skill
-
-这套方法论是从老徐自己 30+ 篇口播稿里提炼、沉淀成的一个**自足 skill**——定位 / 选题 / 成稿 / 诊断 / 去味 全带上，学员只装这一个。里面的母题库、公式、五维诊断、去车轱辘话法，**料全部来自老徐本人的实践**，版权归老徐，随你发给学员、随你商用。
-
-## 标杆样例（按真实播放立标杆，不按稿子好看程度）
-
-`~/口播稿/` 里是老徐 30+ 篇成稿。**自然流量数据标杆**（播放数 asOf 2026-07-31，xuzhaomeng.ai 每日同步）：
-
-- `AlphaCouncil_AI炒股实验_口播稿_v4.md` — **83.1 万播放**（发布题「10万块给AI炒股，16天后大揭秘」）：自问自答推进 + 真金实验 + 每步带费用
-- 「中年辍班，all in AI，一人公司，二月总结」（.txt 原稿）— **57.4 万**：直接报题 + 裸晒收入（3328 元 / 单日 678 / 第二业务线零）
-- `AI算命靠不靠谱_评测台_口播稿_v6.md` — **37.1 万**：成本悬念开场（400 美金 / 2 万次调用）+ 全篇「这就够了吗？还不够」式自问自答
-- 「一人公司都在卖课？普通人如何赚100万？」（.txt 原稿）— **23.1 万**：质疑式开场 + 一道题推演到底 + 结尾对号入座表 + 评论区过招（结构拆解见 `method.md` 推演类骨架）
-
-**商单标杆**：`转转二手MacBook_AI硬件选购_口播稿_v9.md` — 46.6 万（**商单，播放可能含品牌投放，不作纯内容验证**）：先亮答案 + 学员现场 + 对号入座，广告稿怎么写看它和 `formulas.md` 广告桥法。
-
-**爆款共性五条**（从三条自然标杆提炼）：①题材是「我拿真金白银替咱试了」的实测 / 实录，不是旁观评论；②标题和开场带具体数字承诺；③结构都在接观众此刻最想问的——炒股 / 算命全篇自问自答，月报开门就答「股东最关心的钱」；④结论 / 成本前置；⑤裸晒钱数。时长 8–15 分钟都有——长不致命，没实测含量才致命。第 4 名卖课篇补了另一条被验证的路：**纯推演 / 理论题材也能赢**——前提是题目具体到数字、逐点配真案例、结尾能对号入座（案例就是推演稿的"实测含量"）。
-
-`钉钉置身事外_离职叙事_口播稿_v6.md` 是**工艺样例**：格式、逐版修改记录、串联审稿的示范最全，但**从未发布、无市场验证**——破立用推合深稿别当默认答案。带学员时拿爆款拆打法、拿钉钉拆格式——**看方法，不是抄老徐的话题和金句**。
-
-## Source
-
-Rewritten for AICUT from [`chenmisss/laoxu-video-script`](https://github.com/chenmisss/laoxu-video-script)
-(MIT), SKILL.md. The method is that skill's own; the way it runs is AICUT's.