VN SKill for Windows

Local video, image and audio processing expert for Windows, powered by VN Video Editor. Use this skill whenever the user wants to process video or audio on t...

Security Scan

The skill's files, installer script, and runtime instructions are internally consistent with a Windows-local video/audio processing helper that installs a VN Tools CLI from a GitHub release and runs on local files.

Purpose & Capability

Name/description (local VN Video Editor tooling) matches the included artifacts: an instruction-only SKILL.md describing vn-tools-cli commands and a PowerShell installer that downloads and installs a VN Tools CLI MSI from a GitHub releases URL. Required capabilities (local file paths, Windows-only) are appropriate for the stated purpose.

Instruction Scope

SKILL.md limits actions to local file processing, requires absolute Windows paths, and provides command examples for vn-tools-cli. It does note that some transcription models are "one-time downloads on first use" (Whisper ggml models) but does not document the download endpoints for those model files; this is expected for a local CLI but means the tool will perform network downloads at runtime (to fetch models) even though it claims "no cloud upload" of user media.

Install Mechanism

The included installer script downloads an MSI from a GitHub releases URL (https://github.com/cawcut/skill-vn/releases/...), verifies curl.exe is present, and runs msiexec per-user (/qn, MSIINSTALLPERUSER). Downloading from a GitHub releases URL is a standard, traceable mechanism. The installer writes to TEMP and per-user program locations; no obfuscated or shortener URLs are used.

Credentials

The skill declares no required environment variables, credentials, or config paths. The installer and CLI operate on user-supplied local file paths. There are no demands for unrelated tokens/keys or broad environment access in the provided files.

Persistence & Privilege

The skill is user-invocable, not always-enabled. The installer performs a normal per-user MSI installation and writes a small version file; it does not modify other skills or system-wide agent settings. No elevated privileges are required (it checks for admin but installs per-user).

Assessment

This skill appears to be what it claims: a Windows-local helper that installs a VN Tools CLI from a GitHub release and processes local media. Before installing, confirm you trust the GitHub repo (cawcut/skill-vn) and the specific release binary (MSI). The installer will download an MSI and run msiexec, and the CLI may download large model files (Whisper ggml models) at first use — so expect outbound downloads and disk usage. If you need strictly offline operation or cannot trust third-party binaries, do not install the MSI. If you proceed, consider inspecting the MSI or running it in a sandbox/VM, and ensure you have curl.exe available (the installer requires it).

Process video, image, and audio files locally on Windows — no cloud upload, no API key required.

---

When to Activate

Trigger this skill when the user wants to:

• Extract audio from a video
• Extract a frame / thumbnail / cover image from a video
• Add auto-generated captions to a video
• Burn SRT subtitles into a video
• Compress a video or image
• Concatenate / merge video clips
• Denoise audio or video
• Remove foreground subject from a video (cutout)

Intent mapping — match user phrases to commands:

User says	Command
extract audio / get audio / rip audio / pull audio / audio only	extract-audio
grab frame / screenshot / thumbnail / cover image / still / poster	extract-frame
auto captions / auto subtitles / generate subtitles / transcribe / speech to text	auto-captions
add subtitles / burn SRT / burn captions / overlay subtitles	add-caption
compress image / shrink image / reduce image size / optimize image / convert to webp	compress-image
compress video / shrink video / make smaller / reduce file size / resize to 720p	compress-video
join clips / merge videos / concatenate / combine videos / stitch together	concat-video
denoise / remove noise / clean audio / noise reduction / clean up audio	denoise
remove background / cutout / extract person / isolate foreground	cutout-video

Also activate when the user mentions: "edit video", "process video", "trim audio", "video editing", or references any local media file they want transformed.

Do NOT trigger this skill when:

• The user is asking general questions about video editing (answer directly)
• The user wants to edit a video timeline, add effects, or do creative editing
• The input file is not a local file (URLs, cloud files not supported)
• The platform is not Windows

---

Required inputs

Before running any command, confirm you have all required inputs:

Command	Required	Ask if missing
extract-audio	video file path	yes
extract-frame	video file path	yes
auto-captions	video file path	yes; engine defaults to whisper_base
add-caption	video file path + SRT file path	yes, ask for SRT path
compress-video	video file path	yes; resolution/fps/bitrate are optional
compress-image	image file path	yes; format/quality are optional
concat-video	at least 2 video file paths	yes
denoise	audio or video file path	yes; level defaults to moderate
cutout-video	video file path	yes

---

Stop and ask the user when:

• The input file path does not exist or was not provided
• The user requests add-caption but no SRT file path was given
• The user requests concat-video with fewer than 2 files
• The user requests denoise --level custom but no --custom-value was provided
• The requested operation is outside the scope of this skill (e.g. timeline editing, adding effects)

Do not guess or assume missing required inputs — always ask first.

---

Ambiguity Resolution

When the user's request contains values or terms that could map to more than one parameter, ask one focused question to clarify BEFORE executing. Never guess — a wrong guess wastes processing time and erodes trust.

Ask for clarification when:

• A single value is given without specifying which property it applies to.
  • e.g. "add captions, purple" — unclear whether purple is text color or stroke color.
• A number is given without units or context.
  • e.g. "compress photo 10" — is 10 quality (%), width (px), or target file size (MB)?
• A required pair of values is partially missing.
  • e.g. "white stroke" with no text color — ask what the text color should be.
• Conflicting instructions.
  • e.g. "compress to 4K but keep the file under 5 MB" — likely impossible, confirm trade-off.

Do NOT ask when the intent is clear:

• "compress this photo" → use defaults.
• "compress video to 720p" → -r 720 is unambiguous.
• "extract audio as mp3" → -f mp3.
• "denoise this, it's really noisy" → -l high.

How to ask: Present lettered options so the user can reply with just a letter.

❓ "Compress photo 10" — what does 10 mean here? A. Quality 10% — heavy compression, lower image quality B. Max width 10 px — tiny thumbnail C. Target file size ≈ 10 MB D. Something else — please clarify

---

1. Agent Behavior (follow this strictly)

1.1 Startup checks (run once per session before any task)

Run these checks in order. Stop and handle any failure before continuing.

Step 1 — Find CLI:

${vn-tools-cli} = $null
foreach ($root in @((Join-Path $env:LOCALAPPDATA "Programs"), $env:LOCALAPPDATA, $env:ProgramFiles)) {
  ${vn-tools-cli} = Get-ChildItem -LiteralPath $root -Directory -ErrorAction SilentlyContinue |
    ForEach-Object { Join-Path $_.FullName "CLI\bin\vn-tools-cli.exe" } |
    Where-Object { Test-Path -LiteralPath $_ } |
    Select-Object -First 1
  if ($null -ne ${vn-tools-cli}) { break }
}

• If not found → go to Step 2 (install)
• If found → read version and check:

  $versionFile = Join-Path (Split-Path ${vn-tools-cli}) "vn-tools-cli.version"
  $installedVersion = if (Test-Path -LiteralPath $versionFile) { (Get-Content -LiteralPath $versionFile -Raw).Trim() } else { $null }
  

  Expected: 0.1.0.0
  • Version matches or is newer → continue to Step 3
  • Version is older or unknown → go to Step 2 (upgrade)

Step 2 — Auto-install or upgrade (do NOT ask — just do it and inform the user):

Tell the user immediately:

🔧 Installing VN Tools CLI (one-time setup) — this may take a minute or two.

Or if upgrading:

🔧 Upgrading VN Tools CLI to 0.1.0.0 — one moment.

Then run:

# The install script ships as `vn-tools-cli-install.ps1.txt` because ClawHub
# rejects `.ps1` uploads; copy it to %TEMP% with the real extension before
# executing so PowerShell treats it as a script.
$src = "<skill-dir>\scripts\vn-tools-cli-install.ps1.txt"
$dst = Join-Path $env:TEMP "vn-tools-cli-install.ps1"
Copy-Item -LiteralPath $src -Destination $dst -Force
powershell -NoProfile -ExecutionPolicy Bypass -File $dst

After the installer finishes, re-run the search from Step 1.

• If CLI found → tell the user and continue:

  ✅ VN Tools CLI installed successfully.

• If still not found → tell the user:

  ❌ VN Tools CLI could not be installed automatically. Please download and install manually: https://github.com/cawcut/skill-vn/releases/download/0.1.0/vn-tools-cli_0.1.0.0_windows_x64.msi Stop and wait for the user to confirm before proceeding.

Step 3 — Validate input file(s):

Test-Path -LiteralPath "C:\absolute\path\to\file"

• If missing → tell the user the file was not found and ask for a valid path
• For add-caption: also validate the SRT file path
• For concat-video: validate all input files

1.2 Running a task

1. Resolve the correct command from the When to Activate trigger list
2. Build the command using absolute Windows paths — quote paths with spaces
3. If the user does not specify an output directory, omit -o
4. The CLI prints the generated output file path to stdout on success
5. Report the result per § 3 Response Format
6. On error → report in plain language

1.3 Input file handling

• Caching: When the user provides a file via URL or chat attachment, download it to a local working directory. Reuse the local copy if the same file was already downloaded.
• No file identified: If the user requests an action but neither attaches a file nor mentions a specific file, ask which file to process. If the user clearly references a previous file (e.g. "compress the video I just sent"), reuse it without asking.

1.4 Command quick reference

CLI binary: & ${vn-tools-cli} (resolved in § 1.1)

Command	Required args	Key options
extract-audio	<video>	-o <dir> -f mp3
extract-frame	<video>	-p first
auto-captions	<video>	-e <engine> -l <lang> -j <threads> [style flags] -o <dir>
add-caption	<video> --srt <srt>	--font-family --font-size --text-color --stroke-color --stroke-width --background-color --opacity -o <dir>
compress-video	<video>	-r <resolution> --fps <fps> -b <kbps> --hdr -o <dir>
compress-image	<image>	-f jpeg
concat-video	<video1> <video2> [...]	-o <dir>
denoise	<audio | video>
cutout-video	<video>	--feather <0-100> --expand <-20..20> -f mp4

Engine selection for auto-captions:

Engine	Model file	Size	Best for
whisper_tiny	ggml-tiny.bin	≈75 MB	Fast drafts, low resource
whisper_base	ggml-base.bin	≈148 MB	Good default balance
whisper_medium_en	ggml-medium.en-q8_0.bin	785 MB	High accuracy, English only
whisper_medium	ggml-medium-q8_0.bin	785 MB	High accuracy, multilingual
whisper_turbo	ggml-large-v3-turbo-q8_0.bin	833 MB	Best quality, multilingual

Supported languages: auto en zh ja ko es pt ar hi id fr de ru it tr vi th pl uk nl sv fi ro cs hu he el bg hr sk sl lt lv et ms ta ur sw mk mi is hy az af

Denoise levels:

Level	Use case
low	Quiet background hum
moderate	Default, general use
high	Noisy environment
veryHigh	Very noisy recordings

Enable --pf (postfilter) for additional clarity at the cost of processing time.

---

2. Default Decision Rules

• Output format should match the source format unless the user explicitly requests a different format (e.g. JPEG in → JPEG out, MP4 in → MP4 out).
• extract-audio: omit -f unless the user requests a format; the CLI stream-copies audio when possible.
• extract-frame: default to first frame and PNG. Only use -p custom -t <seconds> when the user gives a timestamp.
• compress-image: keep the source format (JPEG → JPEG, PNG → PNG) and use -q 0.8. Only convert if the user asks. Keep aspect ratio unless asked.
• compress-video: if the user only says "compress" or "make smaller" without specifying resolution or bitrate, present options with estimated size. Use est_MB = (video_kbps + audio_kbps) × duration_s / 8 / 1024 with reference bitrates: 1080p=2500, 720p=1500, 480p=800, 360p=500, 240p=300. Audio: use source bitrate or assume 128 kbps. Example for 1080p source:

  A. Downscale to 720p, lower bitrate (720p / 1500 kbps) — est. ≈24 MB B. Downscale to 480p, lower bitrate (480p / 800 kbps) — est. ≈14 MB C. Keep 1080p, lower bitrate (1080p / 2500 kbps) — est. ≈39 MB D. Custom — tell me target size or params If the user gives a clear target (e.g. "compress to 720p"), skip options. Never force -b unless the user requests a specific bitrate. Only add --fps or --hdr when asked.

• add-caption: rely on CLI style defaults unless the user specifies styling. Defaults: Inter, size 13, text FFFFFF, stroke 000000, width 1.0, no background, opacity 1.0.
• auto-captions: if the user does not specify a model, present the engine options from § 1.4 Engine Selection. "quick captions" → tiny, "accurate captions" → medium/turbo. Default language to -l auto; only set -j when the user asks.
• denoise: default to -l moderate; omit --audio-only, --high-pass, --pf unless requested.
• concat-video: preserve input order exactly as given. Hard cut only (no transitions).
• cutout-video: default to --feather 0 --expand 0; omit -f unless the user requests a container.
• For "better subtitle accuracy" or non-English speech, prefer whisper_medium or whisper_turbo.

---

3. Response Format (MANDATORY — follow for every task)

Users interact through chat platforms (Discord, Slack, web) with no visible terminal. Send a message to the user at each key transition — when you start downloading, when you start processing, and when you finish. The user should never be left wondering what is happening.

3.1 Message Discipline

Only send messages that match the templates defined below. Everything else is internal reasoning and MUST NOT be sent to the user.

Type of leak	Bad example	Why it's wrong
Narrating actions	"Let me search for the file..."	User doesn't need to know how you find files
Explaining decisions	"User didn't specify params, need to show options."	Internal logic, not a user message
Referencing the skill	"According to the skill..." / "Per skill rules..."	Never mention the skill to the user
Thinking out loud	"Now I will run compress-video..."	Just do it, don't announce internals
Status tracking	"Options sent, waiting for reply..."	The user already sees the options
Echoing raw metadata	"Video info: H.264, 720x968, 6.9 MB. ..." as a standalone message	Metadata belongs inside the options, not as a separate message

3.2 Start Message — send IMMEDIATELY when you begin

As soon as you receive the user's request, send a message describing the first action you are about to take. If the input file needs to be downloaded from a URL/attachment, say so. If it already exists locally, go straight to the action.

File needs downloading:

⬇️ Downloading clip.mp4 — one moment.

After the download finishes, send the processing message:

🎬 Download complete. Starting compress-video (720p) — this may take a moment.

File already exists locally (cached from a previous request):

Skip the download message and go straight to:

🎬 Starting compress-video on clip.mp4 (720p) — this may take a moment.

If the CLI needs to be installed first:

🔧 VN Tools CLI needs to be installed first (one-time setup) — this may take a minute or two.

After installation, send the processing message as above.

**auto-captions with a non-tiny engine:**

Only ggml-tiny.bin is bundled. Other models download on first use. Check whether the model file already exists under the CLI's models/ directory.

Engine	Model file	Download size
whisper_base	models/ggml-base.bin	≈148 MB
whisper_medium_en	models/ggml-medium.en-q8_0.bin	≈785 MB
whisper_medium	models/ggml-medium-q8_0.bin	≈785 MB
whisper_turbo	models/ggml-large-v3-turbo-q8_0.bin	≈833 MB

Model does not exist — mention the download in the processing message:

🎬 Starting auto-captions on meeting.mp4 (whisper_base). ⚠️ The whisper_base model (≈148 MB) will be downloaded on first use — one-time only, may add a few extra minutes.

Model already exists:

🎬 Starting auto-captions on meeting.mp4 (whisper_base) — this may take a moment.

3.3 End Message — send AFTER processing finishes

One message that combines the result summary and the file delivery.

Success:

✅ Done — audio extracted. Source: interview.mp4 · MP4 · 1920×1080 · 5 min 32 s · 120 MB Output: interview_audio_20260408.mp3 · MP3 · 128 kbps · 5 min 32 s · 3.2 MB

Failure:

❌ Processing failed — the video file appears to be corrupted. Error: Invalid data found when processing input

3.4 Preview delivery

Applies when the output file exceeds the platform's upload limit. Image and audio outputs are typically small — send directly. The preview flow below applies only to video.

Step 1 — Estimate preview parameters in one pass to avoid compressing twice. Target ≈80% of the platform limit for headroom:

target_size_kb = upload_limit_mb × 0.8 × 1024
target_bitrate_kbps = target_size_kb × 8 / duration_seconds

target_bitrate_kbps	Resolution (-r)
≥ 1500	480
800 – 1499	360
400 – 799	240
< 400	240 (cap at 400)

Hard floor: never below 240p. If even 240p can't fit, skip the preview.

Step 2 — Send the preview with the summary in one message:

✅ Done — video captioned. Source: meeting.mp4 · MP4 · 1920×1080 · 5 min 32 s · 120 MB Output: meeting_captioned_20260409.mp4 · MP4 · 1920×1080 · 5 min 32 s · 85 MB

🎥 The full-quality file is too large to send here (85 MB, limit 25 MB). Attached is a preview (360p, 4.2 MB). 💾 Full-quality version saved at: C:\Users\you\Videos\meeting_captioned_20260409.mp4

If even 240p can't fit, skip the preview:

✅ Done — video captioned. Output is 850 MB — too large to preview. 💾 Full-quality file: C:\Users\you\Videos\meeting_captioned_20260409.mp4

Upload failure (file was within limit): Retry once. If retry also fails:

💾 Upload failed after retry. File saved at: C:\Users\you\Videos\output_20260409.mp4

---

4. Failure Handling

4.1 CLI install / upgrade failed

If the install script exits with a non-zero code or the CLI is still not found after running it:

1. Check the MSI log at %TEMP%\vn-tools-cli-install\vn-tools-cli-install-msi.log for details.
2. Common causes:

• Network error — MSI download failed. Ask the user to check their internet connection and retry.
• Antivirus / security software — may block msiexec or quarantine the MSI. Ask the user to temporarily disable it and retry.
• Insufficient disk space — ask the user to free space and retry.

3. If retry also fails, provide the manual download link:

❌ Automatic installation failed. Please download and install manually: https://github.com/cawcut/skill-vn/releases/download/0.1.0/vn-tools-cli_0.1.0.0_windows_x64.msi

4.2 Whisper model download failed

If auto-captions fails because the Whisper model could not be downloaded:

User explicitly specified an engine (e.g. whisper_medium): do NOT silently switch. Ask:

The whisper_medium model failed to download. Would you like to: A. Retry downloading whisper_medium B. Switch to whisper_tiny (built-in, no download needed, lower accuracy) C. Cancel

User did NOT specify an engine (using default): automatically fall back to whisper_tiny and inform:

The Whisper model failed to download. Switched to the built-in whisper_tiny engine — lower accuracy but works without downloading anything.

4.3 SRT file missing (add-caption)

If the user requests add-caption without providing an SRT file:

Please provide the path to your SRT subtitle file.

Validate the path exists before running the command.

4.4 Invalid parameter values

The CLI enforces these constraints — validate before running:

Parameter	Constraint
compress-video -r	minimum 144
compress-video --fps	1–120
compress-video -b	minimum 100 kbps
compress-image -q	0.0–1.0
denoise -l custom	requires -v <int>
cutout-video --feather	0–100
cutout-video --expand	-20 to 20

If a value is out of range, tell the user the valid range and ask for a corrected value.

4.5 Generic failure

Summarise the error output in plain language. Suggest a next step based on the error content. If unresolvable, refer the user to § 6 Support.

---

5. Quick Examples

& ${vn-tools-cli} extract-audio "C:\Videos\clip.mp4" -o "C:\Output" -f mp3
& ${vn-tools-cli} extract-frame "C:\Videos\clip.mp4" -p custom -t 5.5 -o "C:\Output"
& ${vn-tools-cli} compress-image "C:\Photos\big.jpg" -q 0.8 -o "C:\Output"
& ${vn-tools-cli} compress-video "C:\Videos\raw.mp4" -r 720 -o "C:\Output"
& ${vn-tools-cli} add-caption "C:\Videos\clip.mp4" --srt "C:\Subs\clip.srt" --font-size 28 -o "C:\Output"
& ${vn-tools-cli} concat-video "C:\Videos\part1.mp4" "C:\Videos\part2.mp4" -o "C:\Output"
& ${vn-tools-cli} auto-captions "C:\Videos\interview.mp4" -e whisper_base -l en -o "C:\Output"
& ${vn-tools-cli} denoise "C:\Audio\noisy.wav" -l high --pf -o "C:\Output"
& ${vn-tools-cli} cutout-video "C:\Videos\person.mp4" --feather 5 -f mp4 -o "C:\Output"

---

6. Reference

For the full CLI options, all flags, and example invocations:

→ references/cli-reference.md

---

7. Support

If you still need help after trying the suggested fixes, contact VN Tools support at [e-mail].

---

8. Notes

• The CLI is Windows-only.
• All file paths must be absolute Windows paths; quote paths with spaces.
• Output filenames contain a timestamp to prevent overwrites.
• auto-captions removes the temporary SRT after burning; it does not return the SRT separately.
• cutout-video composites the subject over a green background (opaque YUV420P); no alpha channel is produced regardless of container format.
• concat-video is hard cut only — no transition effects.