agentleFS
vargHQvargHQ/skills/stickman-video/SKILL.md

stickman-video

Skill34 starsChanged 8 months agoMITView on GitHub

Make a viral stickman video in code, with no video model: a jointed SVG stick figure (round head, two-bone arms and legs with fixed bone lengths, poses written as joint angles) acts out a survival-choice story, a POV story, a "fake friend vs real friend" contrast, a history explainer or a stick fight, with a DAY counter, rules that appear one at a time, word-timed captions, procedural sound and a two-choice "FOLLOW FOR DAY 2" ending, rendered to a 1080x1920 MP4 with HyperFrames. Ships the working rig (planted-foot walk, breathing idle, head turns, blinks, shock, crouch), script templates with word counts, the fixes for stiff head-only stickman animation, and a 24 s example whose voice cost $0.04 (4 credits). Use when someone asks for a stickman video, stick figure animation, a faceless stickman channel, a survival choice story, POV stickman, or wants to "make videos like the viral stickman reels". For Google Flow or Veo clip prompts use kaomei/stickman-video-director instead.

  • Reads credentials

The file

We split this skill into the rules it gives an agent

12 rules. About 0 are common, 8 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈8

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Ask once, in one message, with a default for each.

    line 44
  2. Distinctive:

    If they bring a premise, use it.

    line 54
  3. Very specific:

    Look-around: over the shoulder (lookX -1), then up, then at the camera on the line that matters.

    line 123
  4. Distinctive:

    Never both arms in unison, never a perfectly mirrored pose.

    line 124
  5. Distinctive:

    No principal picture holds still for more than 2.5 s.

    line 125
  6. Very specific:

    Write the film in index.html:

    line 144
  7. Very specific:

    Look: bunx hyperframes snapshot --at <12 beat times> --no-end, then --range <walk start>,<walk end> --frames 8 for the walk, and zoomed …

    line 150
  8. Distinctive:

    Do not copy a named channel's characters, names, catchphrases, signature outfit or episode premises.

    line 188
  9. Distinctive:

    Do not use real people's names, faces or cloned voices.

    line 189
  10. Distinctive:

    Do not present invented stories as true, and do not invent facts in explainers.

    line 191
  11. Distinctive:

    Do not promise views or income: the numbers in the posts above are the posters' claims.

    line 192
  12. Very specific:

    Do not use Math.random, Date.now, timers or the previous frame as state.

    line 193

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Stickman video: a jointed figure, a few rules, a choice

Stickman reels are a format wave on X, TikTok and Shorts in autumn 2026, and the pitch that travels is "100% code, $0 video generation" (@finnardley, 65,440 views). The complaints are just as consistent: limbs too long, a character that only moves its head, monotone animation (a user replying to the stickman-video-director author). This skill fixes those with a real rig: bones of fixed length, poses as angles, legs solved by inverse kinematics, and several motion layers running at once.

The example in assets/example/ is a 24.3 s "Day 1: lighthouse rules" reel built with this skill: an original premise, a 60-word voice for $0.04 (4 credits), procedural sound, word-timed captions, three rules, a two-choice ending, rendered in 17 s.

1. Ask, then go

Ask once, in one message, with a default for each. Then write the script and show it before building.

AskOptionsDefault
Formatsurvival choice (DAY N + rules + 1/2 ending), POV, contrast (fake vs real), history explainer, stick fightsurvival choice
Length15–25 s (one episode, Shorts/Reels), 45–60 s, longer = a series of episodes20–25 s
Ratio9:16, 16:9 (YouTube long form), 1:19:16, 1080x1920, 30 fps
VoiceAI narrator on varg at about $0.04 (4 credits) per 60 words, their own recording, no voice (captions + sound only)AI narrator, calm and tense
Stylenight set with a white figure; paper white with a black figure; retro TV glow; one set drawn in codenight set

If they bring a premise, use it. If not, offer three one-line premises that are not copies of a known channel's episodes (see section 9).

2. Five formats and why each travels

View counts are the X posts' own counters on the collection date (2026-11-13). Follower and income figures inside those posts are the posters' claims, not verified.

FormatWhat it isWhy it travelsEvidence
Survival choiceSecond person ("you"), a DAY counter, 3–4 rules that appear one at a time, a rule gets tested, the ending card "1: … / 2: …" and "FOLLOW FOR DAY 2"The choice asks for a comment, the DAY counter promises a series, the rules give the viewer something to check against the picture@finnardley code-only prompt with exactly this structure, 65,440 views; @brainextends zombie-on-a-farm interactive episodes ("22M views, 154K followers in 5 days"), 50,885 views; @laurgrowth empty subway, "1: GET ON / 2: STAY", 6,615 views; @CrazyKaomei doomsday-survival style with an A-or-B ending, 21,050 views
POV"POV: you're the last one in …"; the viewer is the figure, the first frame is the situationNo setup needed; the hook is the situation itself; easy to make 4–5 hook variants of one story@FeyiLux "how to start the POV stickman niche" (and: the niche is crowded, write 4–5 hooks), 26,076 views; @nolantoken creepy stickman stories with choice endings, an app account claiming 56.3k followers in 15 posts, 12,034 views
ContrastTwo figures, the same situation twice: fake friend vs real friend, you vs you, white figure vs yellow figure; one line of textReadable with the sound off in under a second; relatable; ends in "tag your real friend"@ghosstty_ on 4 "fake friend vs real friend" reels over 50M views each (top 97.3M), 7,620 views; @maubaron on 11-second stick-figure clips, black background, one line of text, "225M views in a few days", 13,229 views
History explainerA deadpan narrator; a grand event told through one small stickman's absurd problem; each claim one pictureThe contrast between the serious voice and the silly picture; evergreen topics; long watch time@CrazyKaomei "history picture book" style, "deadpan absurd contrast", 57,439 views; @sarpstar an 8-minute documentary made in code overnight, 540,864 views; a reply asking how to get this style with a consistent character was pointed to the stickman-video-director repo
Stick fightTwo figures stand in for two teams, apps or models and fight in three exchangesRivalry, a clear winner, short; the "made 100% in code" claim is part of the post@rayanabdulcader two chat-app dock icons fighting as stick figures, 100% code, 137,490 views; @Framer_X a stickman fight used as the motion reference for other characters, 72,629 views

Also cited for the size of the wave: @Aygramm_ "stickman animations are the new gold rush", 40,656 views; @theoardley listing "stickman survival stories" among free Claude video skills, 499,979 views.

3. Script templates

Pace, measured. ElevenLabs v3 "Adam" read the 60-word example in 25.1 s raw (2.4 words/s, sentence pauses of about 0.4 s). Shortening every pause to 0.2 s gave 22.2 s (2.7 words/s) without stretching speech. Plan at 2.5 words per second of voice, then tighten pauses. Leave 2 s after the last word for the ending card.

Survival choice, 20–25 s, 55–65 words

BeatSecondsWordsLine patternPicture
Hook0–1.22"Day one."DAY N title already on frame 0 over the figure mid-stride
Situation1.2–510–12"You're the new [role]. The last [role] left three rules."walk in, stop, look back over the shoulder, look up, look at the camera on "three rules"
Rule 15–7.66–8"Rule one: [simple instruction]."rule card 1 lands on the word "Rule"; the figure shows it (points at the thing)
Rule 27.6–12.312–14"Rule two: [instruction]. If [condition], [don't do X]."the condition comes true on screen; the figure almost does X and stops
Rule 312.3–15.27–8"Rule three: never [the thing about to happen]."the figure walks to the danger and looks at it
Break15.2–20.212–14"It's [time]. [The rule is tested]. And [a twist that makes it personal]."the HUD clock flips; shock jump; crouch and tremble; the camera creeps toward the danger
Choice20.2–22.38"[Option 1], or [option 2]?""1:" lands on option 1's first word, "2:" on option 2's; the figure looks between the two
Hold22.3–24.30none"FOLLOW FOR DAY 2" for at least 2 s

The example's script: "Day one. You're the new lighthouse keeper. The last keeper left three rules. Rule one: the light never stops turning. Rule two: count the ships. If there's one more than yesterday, don't wave. Rule three: never open the door after midnight. It's twelve-oh-one. Someone knocks. And the voice outside sounds like yours. Open the door, or keep the light on?" Write numbers as words for the voice ("twelve-oh-one") and fix them back in the captions.

POV, 15–20 s, 35–50 words

0–1 s: the text "POV: [situation]" on frame 0 (on screen; the voice may read it). Then three escalating beats of 3–4 s, 8–12 words each, each one a new action or a new thing in the set. End on a question or a choice (6–8 words). Write 4–5 hook variants and test them as separate posts.

Contrast, 10–15 s, 0–25 words

0–1.5 s: one line of setup text ("When you trip in public:"). Then the same staging twice, 4–6 s each, label on top ("FAKE FRIEND" / "REAL FRIEND"): the same trigger, opposite reactions, a 0.4 s hold on each reaction. 1–2 s ending: "Tag your real friend". Often no voice: music bed and two or three effects.

History explainer, 45–60 s, 110–150 words

"In [year], [grand event]." in a deadpan voice, then one claim per sentence, one picture per claim, a change every 2–3 s. Every third beat cuts from the grand scale (a map, a crowd of identical figures) to one small figure's problem. End on an ironic payoff and a question. Every fact needs a source you can name; never invent dates, numbers or quotes.

Stick fight, 10–20 s, 0–10 words

Name tags over two figures (generic stand-ins, not real logos). Three exchanges: anticipation 0.15–0.2 s (wind up the opposite way), hit in 2–3 frames, reaction held 0.3–0.5 s, a hit sound on the contact frame. A reversal in the third exchange, then a winner pose and one line of text.

4. The rig, and the fixes for stiff stickmen

The figure is assets/stickman-rig.html (a 9-second pose tour that renders as is) and the same code as assets/example/rig.js. API, channel table and pose values: references/rig.md.

ComplaintFix in the rig
Limbs too long, proportions driftFixed bones in local units: head radius 36, neck 12, torso 124, upper arm 70, forearm 66, thigh 86, shin 84. Nothing ever stretches; an unreachable foot target stops where the leg ends.
Limbs detach, poses popOne SVG path per limb, recomputed every frame from joint angles. Never swap whole images between poses.
Feet slideThe walk phase comes from the distance the hips have travelled, not from time: the stance foot is planted at any speed. Legs are two-bone IK with the knee bending forward.
Only the head moves; monotoneEvery shot has a body action. Motion is layered: keyed pose, walk cycle, breathing (3.1 s) and weight shift (5.7 s), tremble, blinks, and a line boil (seeded jitter of 0.8 px redrawn 12 times a second).
Everything moves at once, on one curveOverlapping action: the head leads by 0.05 s, the near arm follows by 0.05–0.07 s, the far arm by 0.09–0.11 s. Eases by job: snap for reactions, back for gestures, inOut for looks, walk (short ramps, constant speed) for travel.
The face says nothingA head turn is the eyes sliding across the head (spacing narrows toward profile). Brows appear only with an emotion; an O-mouth for shock; eyes up to 1.6x.
Arms vanish when they cross the bodyThe near arm and leg get a dark outline pass that starts partway down the bone, so the shoulder and hip joins stay clean.

Acting rules:

  • One action per beat: anticipate (0.1–0.2 s), act fast (snap, 0.1–0.15 s), hold the reaction 0.4–0.8 s.
  • Shock is a sequence, not a pose: jump back 40 px with arms up and eyes 1.5x, land with the knees absorbing (hipDrop 34), then crouch (hipDrop 80, lean 26, tremble). Keep the eyes visible after the crouch: drop the near arm to the knees when the figure peeks.
  • Look-around: over the shoulder (lookX -1), then up, then at the camera on the line that matters.
  • Never both arms in unison, never a perfectly mirrored pose.
  • No principal picture holds still for more than 2.5 s. Breathing alone does not count, and neither does a camera move or a text entrance on its own: give the figure an action.
  • Turning around is a 0.16 s flip through edge-on; author a blink on it.

5. Sound and captions

Voice. On varg: check_balance, then search_voices with an intent such as "calm, tense, low male English narrator for a dark survival short", then estimate_cost, then one generate_speech. The example used "Adam" (Deep English Story Voice, varg voice id preset_610bdb815cd5) on eleven_v3: 354 characters, 60 words, $0.04 (4 credits), one take. Download the file, then shorten the pauses: bun tighten-pauses.mjs voice.mp3 assets/audio/voice.wav (pauses over 0.25 s become 0.2 s; speech is never time-stretched).

Word timings, free. bunx hyperframes transcribe assets/audio/voice.wav --model small.en --language en runs whisper.cpp locally (2 s for 22 s of voice) and writes assets/audio/transcript.json. Read it: the example's ASR wrote "twelve-oh-one" as "1201". Do not pay for transcription.

Captions. bun captions.mjs assets/audio/transcript.json <voice start> <words already on a title> <ending card start> writes the groups into index.html: at most 3 words, one line of at most 18 characters (font shrinks from 74 px for long lines), a break at punctuation or a pause of 0.15 s or more, the active word in yellow. Place them under the rules and above the figure's head, inside the 168–1330 px band of a 1080x1920 frame. Hide them while the ending card shows the same words.

Ambience and effects. One bed plus 2–4 effects, each on a visible cause. bun sfx.mjs assets/audio 25 synthesizes everything with ffmpeg, so nothing needs a licence: a sea, wind and lamp-motor bed normalised to -30 LUFS; a low thud for the title stamp; a door knock in three takes at different pitch; an electrical buzz gated by the same on/off table as the lamp flicker in the picture; a whoosh that peaks 0.8 s after it starts. If you use a library sound instead, keep its URL and licence (CC0 or the library's own terms) next to the file. Levels in the example: voice 1.0, bed 1.0, knocks 0.65–0.8, thud 0.45–0.55, buzz 0.3, whoosh 0.4; the mix measured -13.5 LUFS integrated, -1.0 dBTP true peak, no clipping (measured with hyperframes audio, not listened to).

6. Build it in HyperFrames

  1. bunx hyperframes init my-reel --resolution portrait --non-interactive --skip-transcribe (or npx).
  2. Copy assets/example/ into it (index.html, rig.js, captions.mjs, sfx.mjs, tighten-pauses.mjs), or start from assets/stickman-rig.html as index.html for a bare stage. The example has no media: make assets/audio/ with steps 3 and 4.
  3. Voice, pauses, transcript, captions (section 5).
  4. bun sfx.mjs assets/audio <film seconds + 1>; place each effect so its peak lands one frame after its picture.
  5. Write the film in index.html:
    • One clock. A proxy object with a setter, tweened from 0 to the duration on the paused GSAP timeline, redraws the set and the rig for that exact time. Nothing reads a clock, Math.random or the previous frame.
    • The set is SVG inside a camera group (translate · scale · translate from three keyed channels cx, cy, cs). Draw it in code: walls, a window, a door, a lamp. A world wider than the frame lets the camera pan between places.
    • The figure: Stickman.create(group, { scale, ground, blinks }), a Stickman.track(...) of pose keys written as P.move(start, duration, { channel: value }, ease) on the voice's word times, and Stickman.draw(fig, P, t) inside the clock.
    • HUD, rules, captions and the ending card are HTML clips animated on the same timeline. Rules land on the word "Rule"; the clock flips on its word.
  6. bunx hyperframes lint, then bunx hyperframes check --json --no-cache. Use bunx hyperframes, not bunx --bun hyperframes: under the Bun runtime check printed nothing on repeat runs.
  7. Look: bunx hyperframes snapshot --at <12 beat times> --no-end, then --range <walk start>,<walk end> --frames 8 for the walk, and zoomed crops of the figure at the reaction poses.
  8. bunx hyperframes render --quality delivery --output out.mp4, then ffprobe (duration equals the root data-duration), bunx hyperframes audio out.mp4 --json, and a contact sheet decoded from the MP4.

Two traps the example hit:

  • Caption hard-kill. Under a clip host, GSAP records the inherited visibility: hidden as the start value, so tl.set(group, { visibility: "hidden" }, end) leaves every group hidden after a backward seek. Kill caption groups with opacity: 0 only.
  • Two-line captions collide with things drawn in the set (the example's ship numbers). Keep one line per group.

The example is one index.html on purpose (one clock drives the rig and the set), so lint warns that the scenes are not sub-compositions; that warning is expected.

7. Checks before delivery

  • Frame 0 is the hook: the DAY title or the POV line plus the figure mid-action, never black.
  • Limbs attached in every sampled frame, bones the same length, feet on the ground except in a jump, no foot sliding during the walk (8-frame strip).
  • Walk, breathing idle, a head turn, one reaction (shock or crouch) and blinks all appear; nothing principal holds for more than 2.5 s.
  • Rules appear one at a time on their words; the clock or DAY counter changes on its word.
  • Captions one line, word-timed, inside 168–1330 px, nothing readable under the right-hand action rail.
  • The ending: "1:" and "2:" on screen with the spoken options, "FOLLOW FOR DAY 2" for at least 2 s.
  • Sound present, effects on their pictures, about -14 LUFS, true peak at or under -1 dBTP; say it was measured, not heard.
  • check has no errors; duration and size match (1080x1920, 30 fps).

8. Costs

ItemMeasured
Voice, 60 words, ElevenLabs v3 on varg$0.04 (4 credits), one take
Word timings (local whisper.cpp)free, 2 s
Sound effects and ambience (ffmpeg)free
Render, 24.3 s at 1080x1920, 729 framesfree, 17.3 s on an Apple M1 Pro (hardware GPU, screenshot capture)
Total for the example$0.04 (4 credits)

9. Other routes

  • kaomei/stickman-video-director (GitHub, MIT, by @CrazyKaomei, about 5,900 stars in November 2026). Director mode writes a treatment, then one prompt per roughly 10-second clip for Google Flow (Gemini Omni Flash), Veo, Kling, Runway or Luma; you generate and edit the clips yourself. Autopilot renders a narrated, captioned MP4 locally (Node 20+, FFmpeg) after a 20-second sample is approved; the author recommends Codex. Eleven visual styles. Install by pasting the repo URL into your agent with "Install this Skill for me"; the installed skill is called directing-stickman-videos. Use it when the person wants generated clips or one of its styles.
  • Generated 2.5D look on varg. A character sheet and set stills with gpt_image_2_6 (estimate on 2026-11-13: up to $0.06 (6 credits) per 9:16 image), approved first, then animate with kling_v4 (estimate on the same day: up to $0.79 (79 credits) for a 5 s 9:16 clip, priced as text-to-video; price the image-to-video call itself before the batch; kling_v4_flash for drafts). Run estimate_cost before each batch. The honest caveat: video models drift on stick figures. Limbs multiply or thicken, heads detach, hair and clothes appear, and two shots rarely match. Keep shots to 5 s, reuse the same reference, judge every take, and keep anything that must stay jointed (a walk, a fight, a pose that carries the joke) in code. A hybrid works: generated background stills under the code figure.

10. What not to do

  • Do not copy a named channel's characters, names, catchphrases, signature outfit or episode premises. Make your own figure and your own premise; "inspired by a format" is fine, a look-alike is not.
  • Do not use real people's names, faces or cloned voices; use generic stand-ins in fights and contrasts, not real logos.
  • Label AI content where the platform asks for it (TikTok's AI-generated label, YouTube's altered or synthetic content disclosure for realistic material), and write "AI voice" in the caption when the narrator is synthetic.
  • Do not present invented stories as true, and do not invent facts in explainers. Keep survival stories PG: tension, not gore.
  • Do not promise views or income: the numbers in the posts above are the posters' claims.
  • Do not use Math.random, Date.now, timers or the previous frame as state; do not swap images for poses; do not stretch bones; do not put text below 1330 px on 9:16.

Sources

  • X research, collected 2026-11-13: the posts linked in section 2, the stickman-video-director launch and updates by @CrazyKaomei, and user questions under them (limbs and monotone motion, two-bone IK arms and lip sync by loudness).
  • kaomei/stickman-video-director: SKILL.md (Director pacing of about 20–25 voice words per 10 s clip, a visual change every 2–3 s) and references/autopilot.md (no principal-picture inactivity over 2.5 s; camera moves or text entrances alone do not count as animation).
  • HyperFrames CLI 0.9.41 documentation (hyperframes, hyperframes-core, hyperframes-cli skills).
  • Measured on the example: voice job on varg on eleven_v3 for $0.04 (4 credits), whisper.cpp small.en, ffmpeg loudness, HyperFrames render log.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.