What does Auto Tutorial Studio do?
Given a website URL or software name, along with a rough chapter outline, this skill handles:
Drive real operations → Notify user to record → Import captured footage → Build searchable index → Sentence-level aligned editing → Motion effects & transitions → Frame-by-frame validation → Final video
Division of labor: What the skill does vs. what the user does
| Who does it | Description | |
|---|---|---|
| Screen recording | User | User records with OBS / ShareX / system screen recorder or other external tools — the skill does not capture screen. External recording preserves native resolution and clarity, and can capture system audio. |
| Real operations | Skill | The skill performs each step using browser CDP or window automation, recording operation positions. |
| Editing, voiceover, effects, validation | Skill | Fully automated, with safety gates. |
| Footage import | User | After operations, provide the recorded video file path to the skill. |
The skill does not take over the user's hardware mouse: browser operations use CDP element clicks and do not move the real pointer.
Seven golden rules
- Capture is user's responsibility; the skill does not record screen. The skill only handles operations and post-processing. The user must hand over the file after recording. |
- Operations are the skill's responsibility; the recording window is the user's. The skill only starts acting after the user announces "start recording", and immediately notifies the user to stop when done. |
- No console windows may appear on screen during recording. Every command executed by the skill must run without a visible window, and the
guardmust immediately hide any new console windows or overlays — they will be recorded. | - Only hide windows, do not close them. Use
SW_HIDE, restore all withguard stop. | - Each clip must correspond to one spoken sentence. Every beat in
plan.jsondeclares which spoken sentence it serves. Chapter total length must equal the total voiceover length for that chapter. | - Numbers read aloud must be read from the real interface. Vertex counts, prices, version numbers — always from the actual UI. |
- Occlusion audit is a gate, not a report. Before building, perform frame-by-frame OCR on the output window; reject the build if any console / notification / splash screen is found. BGM must make room for vocals. Final output: −16 LUFS, true peak ≤ −1.5 dBFS. |
Phase 0 · Project scoping
Confirm three things, all mandatory:
- Target: Website URL or desktop software name and version.
- Outline: Chapter sequence (the user provides a rough one; the skill fills in details).
- Recording plan: Which screen recording tool the user will use, which display to capture, whether to record system audio, output format and resolution.
Also determine the capture area (full screen or a specific window) and whether to record in segments. Recording one segment per feature is easier to maintain; recording the entire flow in one go is possible but relies on the index for positioning.
Phase 1 · Script writing
build/script.md, each chapter as a ## <heading> with format ## <ChapterID> | <Display Title>:
## ch00 | Introduction
First, look at what I've made.
This figurine was generated from a single image.
## ch01 | Registration and Login
The official website is hi3d.ai, just open it in a browser and you can use it.
- ChapterID: Only lowercase letters, numbers, underscores, and hyphens allowed (required by TTS manifests). Must exactly match the chapter
idinplan.json. Multiple segments within the same chapter are automatically numbered asch01,ch01_1,ch01_2. - Display Title: For chapter cards, can be in the original language.
- ChapterIDs must not be prefixes of each other (e.g.,
ch0andch01would conflict). Recommend usingchNNformat. - Start with the result (the most recognizable visual), then go back to the operations.
- Mention edge-case features in one sentence. Dedicate space to the parameters that determine the outcome and how to verify correctness.
- Save prompts, code, and configuration in fenced code blocks (```) — code blocks will not be spoken, they are purely visual.
- Every spoken segment must have corresponding real footage. Rewrite any segment without footage into something that can be demonstrated visually.
Phase 2 · Voiceover generation
python scripts/studio.py voice --project <project>
Splits by ## <ChapterID> | <Title>, separates by blank lines, skips code blocks and > notes, produces build/audio_map.json, each segment with its measured duration.
Check the output chars/s value: segments below 3.0 indicate TTS anomaly (some synthesis got stretched). Rewrite and regenerate — do not proceed to editing with such segments.
Phase 3 · Real operations and external recording ★
This is the core phase of the skill.
3.1 Prepare recording environment
python scripts/studio.py windows # Find the target window
python scripts/studio.py prep --project <project> \
--keep-title "Chrome" --maximize-title "Chrome" # Hide unrelated windows, maximize target
prep never hides the skill's own terminal (excluded based on ancestor PID).
3.2 Start screen guard ★
This step cannot be skipped. During operations, the skill executes commands (ffmpeg, ffprobe, python subprocesses). Any of these could spawn a console window that lands on the user's capture screen and gets recorded.
python scripts/studio.py guard start --project <project> \
--protect "DeepSeek Harness" --protect "Chrome"
The guard is a background process that scans the window list every 0.25 seconds. It immediately hides (SW_HIDE) any new console-class windows (ConsoleWindowClass, cmd, PowerShell, terminal, xshell, etc.) and input method / notification overlays. It remembers which windows it hid and restores them on guard stop.
--protectprotects applications being demonstrated (matches by title, mandatory).- By default, only hides console windows from the skill's own process tree — does not hide GUI windows. The user's session interface remains visible.
- Use
--hide-own-guito hide GUI windows as well. - The guard's criterion is "real overlap": a window that only overlaps the capture area by a few pixels (e.g., on another monitor) will not be hidden erroneously.
3.3 Notify user to start
Explicitly tell the user: Please start recording. Reply "recorded" when done. Do not act before user confirmation — otherwise preparation steps will be recorded.
3.4 Execute real operations
Proceed chapter by chapter. Each step:
- Browser: Use CDP element clicks, do not move the real mouse.
Record the click coordinates for later callouts:
python scripts/studio.py aim <project> <x> <y> --note "Export format". - Desktop application: Activate window + keyboard-first; use real mouse only when necessary for clicks, but warn the user not to move their mouse during recording.
- When status confirmation is needed, no need to interrupt the user — grab a frame and check yourself:
python scripts/studio.py shot --out <project>/build/check.png
- Login, authorization, CAPTCHA: Keep recording. Let the user handle these; the skill only confirms success or failure. Cut out credentials, CAPTCHAs, keys — you can record before and after login separately, then splice in dependency order.
- Every command executed by the skill must run without a visible window:
scripts/h.ps1starts subprocesses withCreateNoWindow; internal subprocesses always useCREATE_NO_WINDOW.
3.5 Notify user to stop
Immediately after all operations are complete, tell the user: Operations finished. Please stop recording and provide the video file path. Wait for the user to return the path. Do not search for or delete any recording files yourself.
3.6 Cleanup
python scripts/studio.py guard stop --project <project> # Stop guard and restore windows
python scripts/studio.py guard status --project <project> # Confirm no residual console on capture area
guard status lists any console-class windows still visible in the capture area — must be empty to be clean. If any remain, a program is repeatedly popping up windows; have the user re-record that segment.
Phase 4 · Import footage and build index
python scripts/studio.py import --project <project> --file "D:\video\rec.mp4"
python scripts/studio.py index --project <project> --clip rec.mp4 --step 2
index scans the entire video at a fixed step, performing frame-by-frame OCR and recording:
- What text is on screen at each moment (to look up what happens at a given second).
- Visual change points (natural cut points).
- Contamination intervals (console windows, notification bubbles, splash screens — to avoid).
Outputs build/index_<name>.txt (searchable) and .json (structured).
Always read the index before arranging shots — do not guess timestamps.
Phase 5 · Shot planning
build/plan.json:
{
"display": {"region": [0, 0, 1920, 1080]},
"voice": {"voice": "zh-CN-XiaoxiaoNeural", "rate": "+8%"},
"chapters": [
{
"id": "ch02", "num": "02", "title": "Text-to-Image", "sub": "Three-part prompt for fidelity",
"beats": [
{"seg": 0, "src": "rec.mp4", "in": 120, "out": 134},
{"seg": 1, "src": "rec.mp4", "in": 136, "out": 152, "zoom": 1.0}
],
"callouts": [
{"seg": 1, "offset": 1.0, "text": "Three-part prompt\nSubject → Material → Composition",
"x": 120, "y": 680, "pointer": "up"}
]
}
]
}
srcis the filename inraw/(user footage imported byimport).in/outare in seconds.segindicates which spoken sentence this shot serves, counted per chapter starting from 0.zoom> 1 does a push-in (static crop enlargement).calloutsuseseg+offsetto position annotations at the moment they appear, following the voiceover.x/yare annotation positions (can use coordinates recorded byaim).- Chapter
idmust equal the chapter ID inscript.mdand must also equal the segment prefix inaudio_map.json. - Full example in
assets/plan.example.json.
Shot planning rules:
- One shot serves one operation focus; shot length is determined by that spoken sentence.
- The annotation should appear at or before the named UI control; leave reading time for critical feedback.
- For adjacent controls, try to share a single framing rather than frequent jumping.
- Compress real waiting time reasonably, but keep a recognizable "trigger action → see feedback" duration.
- Rewrite any segment without footage into something that can be demonstrated visually, or re-record.
- Avoid contamination intervals marked by the index.
Phase 6 · Editing, validation, delivery
python scripts/studio.py build --project <project>
python scripts/studio.py verify --video <project>/final.mp4 --expect-duration <voiceover total length>
build automates: sentence-timed shot cutting → chapter title cards → annotation overlay → cross-dissolve transition tree → BGM side-chain ducking → loudness normalization.
BGM: Place music file at build/bgm.wav (or build/bgm_final.wav). The builder calculates the ducking curve based on voiceover envelope and mixes it in. Output works without this file, just no background music.
When audit fails: Based on build/audit.txt timestamps, replace the offending window or re-record. Edit plan.json and rerun. Common approaches: shift the contaminated in/out entirely to clean intervals, or ask the user to re-record that segment.
Pre-delivery self-check:
- File plays completely; video and audio tracks are complete; duration close to total voiceover length.
- Frame-by-frame OCR audit passes with zero issues.
- Watch and listen to entire video at normal speed; focus review on: first 3–10 seconds, each feature transition, submission & results, numbers & proper nouns, waiting segments, ending.
- Every explanation point has a corresponding cursor position; no long static image swaps, no empty pauses, BGM does not overpower vocals.
Quick command reference
studio init <project> # Create project skeleton
studio windows # List windows, find operation targets
studio prep --project <project> --maximize-title "Chrome"
studio guard start --project <project> --protect "Chrome" # ★ Keep capture area clean during recording
studio shot --out check.png # Grab a frame to check status
studio aim --project <project> 1001 720 --note "Export format"
studio guard stop --project <project> # Stop guard and restore windows
studio guard status --project <project> # Confirm no residual console windows
studio import --project <project> --file rec.mp4
studio index --project <project> --clip rec.mp4
studio voice --project <project>
studio build --project <project>
studio audit --project <project>
studio verify --video final.mp4
studio status --project <project>
References
references/production.md— Screen recording, voiceover, editing, mixing, acceptance checklistreferences/troubleshooting.md— Symptom → root cause → fixassets/plan.example.json— Full shot plan example