How to use
- Copy the installation command and review the source package before installing it in your own agent environment.
- Check the source document for required tools, credentials and supported hosts.
- Describe your input and desired deliverable; review the output against the source material.
npx skills add https://github.com/mengto/skills --skill making-of-filmBefore you start
- Installation does not include a model subscription, API allowance or permission to access your files.
- Repository stars and installation counts describe activity, not an independent quality or security test.
Rendered from the source repository. Headings, examples, code, tables, links, and referenced images are preserved.
Making-of film
A ~4:30 film in which Meng shows how he made something with AI: the result first, then the workflow as numbered steps, each step his real prompt and its real result, then an honest account of where it falls short and what it cost. It is built the way html-explainer-video builds films (HTML scenes drawn from one clock, every time a word of the voice), scaled up to thirty scenes, dressed in the project's own look, with his digital twin at the key places.
Written for Meng's own setup. This public copy leaves out the IDs of his voice, his twin and his canvases: bring your own cloned voice and HeyGen twin (create-ai-avatar, digital-twin in this repo). Skills named meng-… and youtube-video-release below are his private ones.
Provenance. The Greymourn Castle tutorial ("how Greymourn was made", a browser boss-fight game), 15–19 Nov 2026: drafts 1 → 4g over five days, threads A1 (talking clips), A2 (film and polish), A3 (release), V1 (look and motion), V2 (UI footage), V3 (backgrounds), and 48 of his prompts about the film (164–212 in lothric/qa/PROMPTS.md). Draft 4f had every key frame at 8 or more with two blind judges; 4g put his Kimi twin in the camera at 26 % of the film. The working files are in lothric/docs/tutorial/ of the HTML Pages repo; references/greymourn.md maps them, scene by scene, with the history of what he asked for and rejected. Read it once before the first film you make.
When to use which skill
| The ask | Skill |
|---|---|
| "How I made X", a tutorial or build log of his project, 3–6 min, 16:9, him in it | this one |
| A one-minute explainer of a product or feature | html-explainer-video (this skill's engine is a fork of its engine) |
| A 9:16 reel from a post, a page or a video | html-reel |
| An app window playing live on a web page | html-demo |
| A still, a talking clip or a voice line of him on its own | meng-digital-twin (references/heygen-twin.md there is the twin recipe this film uses) |
| A shot of him inside the game, a three.js or a Blender scene, in costume, saying the film's words (an insert over his camera; Greymourn scene one) | genjutsu-scene (HeyGen Avatar 6 twin to Genjutsu; the camera itself stays the HeyGen twin, rule 7) |
| Export, Studio upload, covers, title A/B test, description | youtube-video-release; covers with his AI face: meng-youtube-cover |
| Scoring rounds to a target ("get it to 9") | workflow-score-to-target (this skill's judging is that, tuned for film frames) |
Meng's standing rules (each one came from a correction)
- Only what's true. Every claim, on screen and in the voice, comes from the project's own record: his prompts,
the git log, scorecards, ledgers, session transcripts. Greymourn's draft 2 had 9 false claims out of 53 (wrong prompt count, "two skills loaded" when Claude loaded one, "29 in the game" when 27 are drawn, a stale score). A wrong number costs a re-voice; check before the voice is made (references/script.md).
- No borrowed names anywhere a viewer can read. Greymourn was modelled on a famous game; the film never names it,
in the voice, on screen, in titles, tags or the description. His prompt 1 is shown with "[…]", and concept images with the game's names in them (a boss bar, stat names, map labels) were blurred. The same goes for any project: no other studio's names, characters or product names on screen unless he says so.
- Price first, then his yes typed in the thread doing the work. Voice (Fil eleven-v4, ~1 credit a line, ~31 a full
script), the twin (HeyGen Avatar 6, $0.10 a second at 1080p), captions, covers: max_credits: 0 or the CLI's cost first, tell him credits and dollars, run on his typed yes. A yes relayed by a coordinator does not count for anything showing his face or voice. Log every spend in the film's ledger.tsv.
- Never take his mouse. Headless Chrome for Testing for every render, capture and probe. No computer use, no
clicks or keys in the Browser pane or any visible window, no screen recording.
- Never change the lighting of his camera. His camera footage and his digital twin's frames are shown exactly as
delivered: no grade, colour matrix, curves, LUT, vignette, brightness, contrast, saturation, relight or sharpening, neither baked into the frames nor applied in CSS or SVG filters over the camera window, and no shade or gradient laid over his face. This holds even when blind judges score a graded version higher (a warm grade won +0.5 with both judges and he rejected it: prompt 212, "I don't want you to touch the lighting at all"; then, 19 Nov, "Make sure you do not change the lighting of the camera"). Only framing may change (position and scale to keep his face on the layout), and the film's grades for plates and captures must be scoped so they never reach the camera.
- The camera at key places only, about 25 % (prompt 202: "only put cameras at key places 25% of the video"):
the intro, each step's title, the verdict, what I'd keep, the end card. Hidden everywhere else.
- His twin, not a model's guess at him. The camera shows his trained HeyGen twin (the Kimi look on Avatar 6),
lip-synced to the film's own voice. He rejected Intone ("I don't speak like that!"), dropped Genjutsu and Seedance clips for this, and kept the twin although blind judges preferred his still (prompt 209: "Keep the twin"). Don't swap the route without asking.
- 4K, single frames, before then after. Every picture you send him is a full 3840 × 2160 JPEG q95, one frame per
file, …-before.jpg and …-after.jpg, never grids or side-by-sides. Video previews 4K60, 20 Mbps or more.
- Shared checkout manners. Commit only your own paths (
git commit -- <paths>), nevergit add -A, stash,
reset, checkout or switch branch; heavy files (frames, clips, renders) live outside the repo in ~/Library/Caches/<project>/. Never call Fil set_visibility.
The film's shape
Thirty scenes, eleven chapters, 4:30. Times are Greymourn's (720 words, 2.7 words a second); a new film keeps the proportions. Full beat list, the per-step loop and the script patterns: references/structure.md.
| Chapter | Time | What it does | On screen |
|---|---|---|---|
| Cold open | 0:00–0:10 | the result, no voice | 4–5 cuts of real footage, 2 s each, the project's own sound, a small badge |
| Hook | 0:10–0:30 | the claim, in numbers, and the honest twist | "Made in three days" over the best footage; then him full frame → half; "176" counts up; his real prompt types out |
| What you get | 0:30–0:49 | the promise, with chapter times | a row per chapter, each lit on its word, its time beside it |
| The stack | 0:49–1:05 | the tools, one job each | one card per tool, landing on its name, lit while said |
| Steps 1–5 | 1:05–3:44 | the workflow, 25–45 s a step | step title (him half frame), then the loop: his real prompt → the real result → what matters marked → the number |
| Where it falls short | 3:44–4:09 | honest limits, with the real scores | "Not at nine", the judges' bars, load time, rejections, "this footage is a scripted bot" |
| What it cost | inside the above | real ledger rows | credits by line, the usage meter |
| What I'd keep | 4:09–4:20 | the workflow in four lines | him half frame, four rows |
| End card | 4:20–4:30 | where everything is | the canvas, big addresses, "Go build your own", him in the corner |
The step loop (from the reference video's format, adapted): say the step → show the real prompt, typed → show the real result → mark the one thing that matters (a ring, a lit phrase, a wipe, a count) → one line that lands it. The most surprising step gets the most time (Greymourn: blind judges, 47 s).
Workflow
Ask at most one question before starting, and only if the project is unclear. Say the defaults you took.
- Read what was built (
references/script.md). The repo, the README,qa/PROMPTS.md(his prompts are the
story), scorecards and judges' verdicts, ledgers, hand-over notes, the git log, the first session's transcript. Write 00-facts.md: every fact the film may state, with its source. If there's only a brief, the facts are the brief's and the film waits for the real captures.
- Research the format (
references/structure.md): one top video in the genre read shot by shot (from its
transcript and frames played headless: never download it), his own channel's numbers, the genre's best 10. Take the format, never the content, look or wording.
- Write the script in his voice (
references/script.md): ~700 words for 4:30, his real prompts quoted exactly,
numbers said when shown, one honest exception, "the trick I'd steal first". Under each line: what the picture does and the word it lands on. Mark lines you wrote for him with ✎. Check every claim against 00-facts.md.
- Set the look from the project (
references/look.md): its palette, its frame motif and metal, its UI language,
in film tokens; the Geist trio for type; real captures of the project at 4K60, frame-exact; backgrounds made from the project's own data and renders. Nothing invented to look like the project.
- Build on a free clock. `html-explainer-video/scripts/words.py --draft script.txt --out film/words --start 10.1
--pace 2.7 (the cold open is the first 10 s). Copy assets/starter/ into the project as film/, with scripts/explainer.js, explainer.css and motion.js beside it. Build all the scenes, one <section> each, from references/motion.md. Look at stills as you go (scripts/render.mjs film.html --stills … --out …`).
- His voice (
references/voice.md): price, his yes, one take of the whole script in his instant clone
("Meng Voice 2026 Instant" on eleven-v4 with [calm]). Word times by scripts/align-voice.py, then scripts/realign-phrases.py phrase by phrase (one-pass whisper put 34 of 89 phrases up to 1.3 s early). Cuts, pauses and re-voiced lines with scripts/voice-edit.mjs, so no scene is renamed.
- The camera and his twin (
references/camera-and-twin.md): the camera plan at key places (~25 %), layouts on
the content's grid, then one HeyGen render of exactly those stretches of the film's voice; scripts/sync-check.mjs proves 0 ms sync; assets/twin.js plays the frames as rendered.
- Render (
references/render.md): every scene at 3840 × 2160, 60 fps, one MP4 each, under the GPU lock
(scripts/render-scenes.sh), then the whole film with sound (scripts/preview.sh): −14 LUFS, true peak under −2.
- Watch it and judge it (
references/judging.md): a frame every 0.5 s in sheets, probes for anything under the
camera or overflowing, every word landing; then 13 key frames, before and after at 4K, two fresh blind judges, the lower score counts, keep only wins and ties. Repeat until the judges plateau (Greymourn: all 8s, none 9).
- Assemble in Fil and hand over (
references/render.md): scenes, voice, the cold open's sound, a music bed,
chapters from scripts/film-facts.mjs; export needs his open Fil window. Then the release package (youtube-video-release, covers with meng-youtube-cover). Nothing goes public: that's his step.
Send him 4K before/after stills as each round lands, and the preview MP4 to ~/Downloads when he asks for it.
Rules, each with the failure it prevents
- Times are words, by name.
T.screenshot, never131.7. A new voice, a re-voiced line or a realignment moved
every scene in Greymourn three times; named times made each a re-render, not a rebuild.
- Real material only, at its real size. Gameplay, UI and the project's world come from captures of the project
itself (frame-exact, 4K60). A drawn window holds the real words. Soft captures were the one thing judges named in every round that layout couldn't fix: capture at 4K or show it smaller.
- The film's UI comes from the project's UI. Greymourn's gilt hairlines, diamond corners and ivory-on-near-black
came from the game's own menus; its backgrounds were the game's castle and level plan. Judges marked down every "generic" device (a Tron grid floor, rainbow foil, god-ray streaks, wireframe clip art) until it was replaced by the project's own.
- Something lands on nearly every phrase, on its word, then holds. Draft 2's faults were mostly empty frames
(a headline alone for 3–5 s, an empty prompt box, a prompt readable for 0.3 s before the cut). Land the first card on the first name, type the prompt while it's said, hold the whole prompt 1.5 s before the cut.
- Calm motion. Expo-out arrivals, soft focus resolving, panels settling flat, a 0.6 px/s drift, a 0.3 s exit
into blur. No shines, glints, sweeps or streaks (his taste across every launch film).
- Everything on the grid. 120 px content margins; the camera on the same margins (corner 300 px circle at
x 1500–1800, y 700–1000 on the 1920 stage; half 600 × 864 at the right). Judges cost the camera a point when it sat 40 px off the edge against 120 px margins, and found every dead band beside it.
- Where the camera leaves, the layout takes the room but keep only reflows that win: of seven tried, three won.
- The honest part is a chapter, not a footnote. Not at nine, the real scores, the load time, what he rejected,
"the gameplay is a scripted bot". It's what makes the rest believable.
- One spend at a time, priced, logged. Everything except voice, twin, captions and covers is free.
Numbers from Greymourn
| What | Value |
|---|---|
| Length | 270.65 s (4:30), 30 scenes (2.8–18.9 s), 11 chapters, every chapter 10 s or more |
| Script | 720 words written, ~690 spoken after cuts; 2.7 words a second |
| Stage | 1920 × 1080, rendered at 2× = 3840 × 2160, 60 fps, 16,237 frames |
| Type | display 150–170, head 84, rows 50–56, body 32, labels 22–24, nothing under 20 (stage px) |
| Camera | 24.5 % of the film on screen by plan (26.7 % with glides); half 600 × 864; corner 300 circle |
| Twin | one HeyGen Avatar 6 render of 8 stretches, 70.8 s, $7.30, 1080p 25 fps, sync 0 ms |
| Voice | his instant clone, 31 credits for the script, 1 credit for two re-voiced lines |
| Render | ~70 min for all 30 scenes at 4K60 under the lock; backgrounds 3–7 ms a 4K frame |
| Sound | −14.4 LUFS integrated, LRA 3.4, peak −1.5 dB; music bed at 0.16 under the voice |
| Judges | 13 key frames; draft 2 at 4–7 → draft 3 at 7–8 → 4f all 8+ (lower of two Opus xhigh judges) |
| Total spend | 32 Fil credits + 197 Higgsfield (tests he asked for) + $7.30 HeyGen |
Verify before you report
- [ ] Every claim on screen and in the voice is in
00-facts.mdwith a source; no borrowed name anywhere visible. - [ ]
scripts/film-facts.mjs film.html: chapters start at 0:00, each ≥ 10 s; the camera is on screen ~25 %. - [ ] Every word lands: probes say nothing runs under the camera or overflows; a frame every 0.5 s looked at.
- [ ] His twin's frames are untouched and
scripts/sync-check.mjsreads 0 ms on every stretch. - [ ] Every kept change won or tied with two blind judges at 4K; scores reported with the lower judge.
- [ ] The preview is 3840 × 2160, 60 fps, −14 LUFS ± 1, true peak under −1 dBTP, AAC 48 kHz.
- [ ] Spend logged in
ledger.tsv, each with his typed yes; nothing public, noset_visibility.
Files
references/structure.md: the beats, the step loop, chapters, research and the reference beat sheet, titles.references/script.md: reading the project, the facts file, writing in his voice, the claims check, the full
Greymourn script as a worked example.
references/look.md: adapting the film's look and UI to the project; captures, plates and backgrounds.references/motion.md: the motion kit (scripts/motion.js), transitions, scene patterns.references/camera-and-twin.md: the camera plan, layouts, and his HeyGen twin end to end.references/voice.md: his voice, word times, realignment, cuts and re-voiced lines.references/render.md: rendering, the preview mix, the watch-through, Fil assembly, the release hand-off.references/judging.md: key frames, blind pairs, the judge prompt, what plateaus at 8 and why.references/greymourn.md: the worked example: where every file is, the 30 scenes, the drafts, his prompts.scripts/explainer.js,explainer.css: the engine, forked fromhtml-explainer-video(half layout, fade to and
from hidden, CAM_DRAW for the twin, under/over hooks, draws that wait for pictures, a focus motion).
scripts/motion.js: word reveals, AR glass panels, gilt linework, highlights, wipes, counts, frame-folder footage,
montage, step titles, lit cards, the scene drift and exit.
scripts/render.mjs,render-scenes.sh,locked.sh,preview.sh: stills, scenes, the GPU lock, the mix.scripts/align-voice.py,realign-phrases.py,fix-onsets.mjs,voice-edit.mjs: the voice's clock.scripts/sync-check.mjs,rects.mjs,film-facts.mjs,judge-pack.py: checks and judging.assets/starter/: an eight-scene film with the structure, the look tokens, the camera plan and the twin hook
wired; assets/twin.js: Greymourn's twin file; assets/judge-prompt.md: the blind judge's brief.