MiniMax-AI/mcode-stack / stacks/jiachen/skills/making-content
making-content
VerifiedContent production in three modes: rewrite, Xiaohongshu (小红书) image post, short video. Rewrite (default): turns an AI-sounding Chinese or mixed Chinese-English draft (post, newsletter, engineering blog, commentary) into the author's restrained style: few emoji, short paragraphs, hedged claims, neutral wording, English terms split by spaces. Triggers: 按我的风格改, 去 AI 味, 润色, 改改这篇, or a vague 帮我改改 when the context is a structured draft. Also for a Xiaohongshu image post from scratch (title, cover, cards, body, hooks); read references/xhs-post.md first. Also turns a document, research, notes or an article into a 60-180 s vertical video (抖音, 小红书, 视频号, B站: real sources, voice-over, captions, cover). Triggers: 做成视频, 做个短视频, 剪个 90 秒版本, 配音, 换成我的声音, or complaints of 字太密, AI 味太重, 容易走神; read references/video-pipeline.md first.
| File | Kind | Size |
|---|---|---|
| .gitignore | other | 167 B |
| SKILL.md | skill md | 25.6 KB |
| evals/evals.json | config | 4.5 KB |
| references/layouts.md | markdown | 8.5 KB |
| references/text-density-lint.md | markdown | 4.2 KB |
| references/video-pipeline.md | markdown | 12.5 KB |
| references/voice-and-music.md | markdown | 5.5 KB |
| references/xhs-post.md | markdown | 9.0 KB |
| requirements.txt | manifest | 233 B |
| scripts/capture.py | script | 6.4 KB |
| scripts/compositor.html | other | 27.8 KB |
| scripts/fetch_assets.py | script | 4.0 KB |
| scripts/lint.py | script | 8.6 KB |
| scripts/render.py | script | 4.6 KB |
| scripts/timeline.py | script | 8.7 KB |
| scripts/tts.py | script | 5.5 KB |
| templates/cover.html | other | 2.9 KB |
| templates/script.example.json | config | 5.0 KB |
| templates/sources.example.json | config | 58 B |
references/video-pipeline.md
View on GitHub# Making short videos
Turns source material into a 60–180 s vertical explainer that reads as hand-edited rather than generated. The pipeline is deterministic: `script.json` → one TTS clip per line with word timestamps → `timeline.json` + `mix.wav` → an HTML compositor that Playwright renders frame by frame → ffmpeg. Each default below fixed a real viewer complaint; the table at the end says which. The voice rules live in this skill's `SKILL.md`.
## Contents
- Setup
- Workflow: 1. Decide with the user first · 2. Write script.json · 3. Fetch assets · 4. Voice · 5. Timeline and mix · 6. Review stills and lint until clean · 7. Render and deliver
- Why the defaults are what they are
- Files
## Setup
Once per machine:
```bash
pip install -r requirements.txt # in this skill's directory
python3 -m playwright install chromium
# also needed: ffmpeg + ffprobe on PATH, and a CJK system font (PingFang SC on macOS, Noto Sans CJK SC on Linux)
```
If this skill's directory has no `scripts/` (the copy is partial), the full package is in the repository the skill was installed from: ask the user to attach it or install the skill again.
One directory per video; run every script from it. `S` is this skill's `scripts/` directory (this file lives in `references/`, so from here `../scripts`).
```text
my-video/
script.json narration, captions, shots (written by you, from templates/script.example.json)
sources.json {"capture key": "https://..."} for real-page captures
shots/ tts/ logos/ fonts/ music/ fetched or generated
compositor.html timeline.json mix.wav stills/ sheet.jpg final.mp4 cover.png
```
## Workflow
Copy into the task list and tick off:
```text
- [ ] 0. Check the topic can win (SKILL.md); stop if it is a dead end
- [ ] 1. Decide with the user: platform + length, voice, hook, music, what to keep out
- [ ] 2. Write script.json and sources.json
- [ ] 3. Fetch assets: fonts, logos, music, page captures
- [ ] 4. Voice: Edge draft now, cloned voice once the user's key is ready
- [ ] 5. Build timeline.json + mix.wav
- [ ] 6. Review stills and lint until 0 issues
- [ ] 7. Render, check, deliver video + cover + description
```
Review in order of cost: script text, then one voice sample, then the contact sheet, then a full render. Send each stage to the user while they are watching, so a wrong direction costs minutes rather than a render.
### 1. Decide with the user first
The user's taste moves the result more than any script, so ask in one round (AskUserQuestion where the harness has it), recommended option first:
| Question | Default |
|---|---|
| Platform and length | 抖音: one 75–95 s cut. 小红书, 视频号, B站: 2–3 min works. A long and a short cut can share one project. |
| Voice | Edge draft first; the user's cloned voice (Volcengine) for the posted version. |
| Hook | 3–4 hooks written from the material: counter-intuitive fact, shocking number, question, conclusion first. |
| Music | One incompetech track (CC BY, credited in the description); offer 2–3 that fit the tone. |
| Keep out | What is private: employer, colleagues, internal project names, personal details. Strip them before writing anything else. |
Put everything only the user can do into this first message, not at the end: creating a voice clone and API key in the Volcengine console (steps in `references/voice-and-music.md`), choosing music, supplying private logos or footage. Build with the Edge voice meanwhile.
Later, stop and ask whenever a choice changes the result a lot (voice, hook, cutting a section); decide smaller things yourself and say what you decided.
### 2. Write script.json
Read `references/layouts.md` for the schema and the ten layouts, and start from `templates/script.example.json`. Editorial rules:
- **Hook within 3 s.** Line 0 says the hook as a question; `meta.hook` shows it as a big overlay for the first 1.6 s. The first shot is the most striking real evidence, such as the source page saying "$0 in revenue".
- **Conclusion by about 15 s.** Line 2 or 3 starts with 先说结论 and states it in one short sentence; everything after is evidence.
- **Spoken, not written.** Short sentences, spoken connectives (你想想, 结果, 那 Cloudflare 呢?, 换我也得买), no parentheses, no enumerations. Say every `n` in your head.
- **Budget.** The cloned voice at speech_rate 25 averages about 5 字/s including pauses (Latin letters count half): 90 s ≈ 430–450 字 of narration, 3 min ≈ 850.
- **`n` and `cap` differ on purpose.** `n` is what the voice says (numbers as words when a voice misreads digits: 十亿美元). `cap` is what the caption shows: digits (10亿美元), identifiers in real spelling (fake_value_), and `/` wherever a caption should break.
- **New picture every 3–6 s, layouts rotating.** A back-to-back repeat is fine for a pair (two cards of a two-part conclusion), not for a whole section. A line can carry two shots; the second cuts on its `anchor` word.
- **Chinese first.** Every English quote gets a Chinese translation set larger than the original.
- **End on the conclusion** plus one question that invites comments (这是我的判断。你觉得最后谁会赢?评论区聊聊). No 下期预告 or 关注我 ending unless the user asks for one.
### 3. Fetch assets
```bash
python3 $S/fetch_assets.py fonts
python3 $S/fetch_assets.py logo oven-sh --repo oven-sh/bun # GitHub owner avatar -> logos/oven-sh.png
python3 $S/fetch_assets.py logo ferris --url https://rustacean.net/assets/rustacean-flat-happy.png
python3 $S/fetch_assets.py music "Digital Lemonade" # also prints the credit line
python3 $S/capture.py --script script.json --sources sources.json --out shots
```
`capture.py` exits non-zero if a phrase is not on the page: copy the phrase from the page exactly, or switch that shot to a `web` layout. If a sandbox allowlist blocks a host, ask the user to drop the file into place. More capture and logo tips are in references/layouts.md.
### 4. Voice
```bash
python3 $S/tts.py --engine edge --out tts-edge # draft: free, no key
VOLC_TTS_API_KEY=... VOLC_TTS_SPEAKER=S_... python3 $S/tts.py --engine volc --out tts-volc # posted version
```
- The key lives in environment variables only, never in a script, file or commit. Suggest revoking a temporary key when the video is done.
- Before voicing everything, voice one representative line (with a number and an English name in it) in 2–3 variants, send them, and let the user pick by ear. The defaults in `tts.py` (speech_rate 25 plus the conversational instruction) are a good starting point.
- Give every voice version its own `--out` directory. A re-run with `--force` or `--only` overwrites clips the user may already have approved.
- `tts.py` speaks `_` and `-` between letters or digits as a space (rusty_v8 → "rusty v8"); listen for other symbols and respell them in `n`.
- Behind a TLS-inspecting proxy, Edge needs `SSL_CERT_FILE` pointed at the proxy's CA bundle.
Console steps, parameters and error codes: `references/voice-and-music.md`.
### 5. Timeline and mix
```bash
python3 $S/timeline.py --script script.json --tts tts-volc --music music/Digital_Lemonade.mp3
```
Writes `timeline.json` (shots, caption cues, chapters) and `mix.wav` (voice and music ducked under speech; no sound effects). It stops with `anchor ... not found in narration` when an `anchor` is not a literal substring of that line's `n`.
A second cut in the same project shares captures, logos and music; give its script, voice clips and outputs their own names and pass them to every later step:
```bash
python3 $S/timeline.py --script script-90s.json --tts tts-90s --music music/Digital_Lemonade.mp3 --out timeline-90s.json --mix mix-90s.wav
python3 $S/lint.py --timeline timeline-90s.json
python3 $S/render.py --timeline timeline-90s.json --audio mix-90s.wav --out final-90s.mp4
```
### 6. Review stills and lint until clean
```bash
python3 $S/render.py --stills # settled end state of every shot -> stills/NN.png + sheet.jpg
python3 $S/lint.py # exits 1 while anything fails
```
1. Open `sheet.jpg` with the Read tool and look: one focal point per frame, nothing clipped, no run of one layout, Chinese larger than English. Open single stills for detail; `render.py --stills 0.5 12.3` grabs exact times.
2. Fix every lint issue with the table in `references/text-density-lint.md` — usually shorter text, a second shot, or context dimmed to texture — and re-run both commands until lint prints `0 issue(s)`.
3. If a frame feels empty after thinning the text, add motion (logo stickers, a counter, bars), not words.
### 7. Render and deliver
```bash
python3 $S/render.py --out final.mp4
cp $S/../templates/cover.html . && python3 $S/render.py --cover cover.html --out cover.png
ffprobe -v error -show_entries format=duration,size -of csv=p=0 final.mp4 # planned length, under 30 MB
ffmpeg -i final.mp4 -af ebur128 -f null - 2>&1 | grep -A1 'Integrated' # -16 to -14 LUFS (target -14)
```
A 90 s render takes about 3 minutes, so give it a long timeout or run it in the background; it comes out around 10–20 MB. Edit the cover's text and logos before rendering it; profile grids crop covers to 3:4, so keep its content inside y 240–1680.
Send `final.mp4` and `cover.png` to the user (SendUserFile where available), and put the post description in the reply:
- the conclusion in about 400 字, as numbered points, for people who will not watch;
- where the facts come from (official blogs, repos);
- the music credit printed by `fetch_assets.py`, e.g. `Music: "Digital Lemonade" Kevin MacLeod (incompetech.com), CC BY 4.0`.
Write the post description with the voice rules in this skill's `SKILL.md` (rewrite mode) before sending it.
## Why the defaults are what they are
Each row is a complaint from a real review of this pipeline and the default that fixed it. Keep them unless the user asks otherwise.
| Viewer said | Default |
|---|---|
| AI 味很重 | Mostly the voice: cloned voice, faster, conversational. Then real sources instead of slides, spoken copy, a cut every 3–6 s. |
| 容易走神 | Hook overlay on the first frame, hook question as line 0, music bed, chapter progress bar. |
| Screenshots too dense, no focal point | `spot` zooms one phrase and dims the rest; dense pages are re-typeset as `web`. |
| One layout for two minutes | Ten layouts, rotated. |
| Keyword sticker too close to captions | Stickers sit in the top zone (y ≈ 330); captions at y 1420. |
| Wants more real-world bits | Logo tiles, mascots and drifting shapes, all animated. |
| rusty_v8 read as "rusty 杠 v8" | `_` and `-` between letters are spoken as spaces. |
| A "biu" sound on every pop-in, then the whoosh felt odd too | No sound effects: `WHOOSH_GAIN = 0`. |
| The whole frame bouncing on the beat looked ugly | No whole-frame periodic zoom or shake; only elements inside a shot move. |
| Half the viewers gone by 20 s: the opening re-showed numbers and narrated how the material was read | The first frame is content, not a still cover; no "we read X" process lines. |
| The hook confused people who had not seen the earlier episode | The first 15 s make sense to a stranger, in units a stranger knows. |
| Too much English on screen | Chinese translation large, original small. |
| 字太密,眼花 | Lint gate: at most 4 text blocks and 36 字 per screen; captions one line of at most 16 字, low on screen. |
| Too empty after cutting text | Ambient motion instead of more words. |
| 下期预告 felt superfluous | No teaser ending. |
| Is 3 minutes right for 抖音? | 90 s main cut; the long cut goes to 小红书 or B站. |
## Files
| File | Use |
|---|---|
| `scripts/capture.py` | Run: real-page screenshots with the quoted phrase located |
| `scripts/tts.py` | Run: one voice clip per line with word timestamps (Edge or Volcengine) |
| `scripts/timeline.py` | Run: `timeline.json` + `mix.wav` |
| `scripts/render.py` | Run: stills and contact sheet, final mp4, cover png |
| `scripts/lint.py` | Run: text-density lint on the rendered frames |
| `scripts/fetch_assets.py` | Run: fonts, logos, music |
| `scripts/compositor.html` | Read when adding or changing a layout. The first `render.py`/`lint.py` run copies it into the project; edit that copy for one video, the skill's copy for all of them. |
| `templates/script.example.json` | Copy: one example of every layout |
| `templates/cover.html` | Copy and edit: the cover |
| `references/layouts.md` | Read while writing `script.json` |
| `references/voice-and-music.md` | Read when setting up the cloned voice, hitting a TTS error, or choosing music |
| `references/text-density-lint.md` | Read when lint fails |
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deployment | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved. | Stars: 197k | Updated | Claude CodeCodex | Verified | Install |
| shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |