Five Prompts, Zero Code: Running a Stick-Figure Video Skill in WorkBuddy to Three MP4s
A no-hype Agent Skill hands-on test: stickman-video-director is not one long prompt, it's a process with acceptance criteria — SKILL.md plus 52 contracts and 41 scripts, each contract owning exactly one thing (style catalog, sound continuity, storyboard templates, acceptance checklist). I installed it in WorkBuddy and ran A director mode: proposal → storyboard → 3 self-contained prompts → local rendering. The part most worth reading is the engineering solution for the "character anchor": instead of writing "keep the character consistent", it freezes the drawing rules word-for-word, specifically to stop the AI from drawing an extra hand. This article walks through the four steps and includes 6 complete copy-paste texts, rendering script ideas, and ffprobe verification results.
I am Xiaohu, a Zhejiang University computer science bachelor's and master's degree holder, a "Po" (breakthrough) operator, author of "AI Super Individual: The Complete Guide to a One-Person Company", focused on AI training, super-individual building, and company efficiency coaching.
Many friends who want to make science content get stuck at the same place: they understand the idea but can't produce the visuals. Hiring someone to make an animated short costs thousands of yuan, and learning editing yourself is too heavy.
Hello everyone, I'm Xiaohu.
Today, using 5 prompts, without writing a single line of code, I'll take you through using WorkBuddy to run an open-source stick-figure video skill from installation to final output — ending with three MP4s you can post directly, plus a reusable rendering script.
Zero cost, and the material never leaves your own computer.
First, the star of this show. White-background, black-line stick-figure science videos have been trending for a while: one little person, a few lines, explaining something hard to describe. No real person on camera; the information density of a piece rides entirely on the visual rhythm.
▲ Searching "Stick Figure VS Animator" on Bilibili sorted by plays: the top five are all stick-figure science animations, the single highest has 10.73 million plays.
I came across an open-source skill, directing-stickman-videos (stick-figure video director). Give it a topic, and it writes out the storyboard, narration, character drawing method, and camera rhythm.
A storyboard isn't the end — from prompts to a postable video, there's still a stretch of road. This time I ran the whole road: installing the skill, generating a proposal, writing prompts, and rendering to video, with real operations at every step and no missing screenshots.
Look at the finished pieces first. The first is a minimal white-background black-line style; the second is a tech style with a red hat — same skill, two style tiers:
Step 1: Install the skill
The skill is essentially a folder: a SKILL.md instruction sheet plus a bunch of contract files. I started a new task in WorkBuddy and asked it to install in one sentence:
▲ ① The instruction is one sentence; ② Press Enter to send
Help me install an open-source skill: stickman-video-director (stick-figure video director). Repository github.com/kaomei/stickman-video-director, install the skills/directing-stickman-videos part completely into WorkBuddy's skill directory. After installation, tell me: what this skill does, what two usage modes it has, and what I should say first.
The installation report was thorough; it laid out the whole inventory first:
▲ ① Two paths laid out as a comparison table; ② It even gave what to say first.
Measured result: SKILL.md plus 52 contracts in references/, 41 scripts in scripts/, and assets and examples are all there.
Each of the 52 contracts handles exactly one thing: style catalog, sound continuity, storyboard template, acceptance checklist.
This skill is not one long prompt; it's a process with acceptance criteria.
Two modes.
A director mode: produces "proposal + storyboard + a prompt per 10 seconds", with human confirmation at every step.
B fully automatic mode: one chain, local rendering straight to MP4. This time I ran entirely in A mode with the default model.
Step 2: One sentence outputs an entire director's proposal
First-time use gets stuck most easily on "where do I start?". It thinks for you; the first sentence is just copy the template:
Use A director mode: topic "Why the busier you are, the less you improve", landscape 16:9, 30 seconds, Style 1 white-background minimal drawing. Give me the complete director's proposal: English narration, storyboard script, character anchor, and sound profile. Stop and wait for my confirmation first.
▲ ① My one-sentence instruction; ② Its reply first states "stop and wait for your confirmation".
The proposal is thick: an opening hook, a five-beat emotional arc, the meaning of each of three accent colors, BGM direction, and English narration split into three segments at 20–25 words per 10 seconds.
To be honest, my biggest worry was character consistency — when AI draws the same person across segments, it often drifts.
Its solution is this character anchor: not a sentence like "keep the character consistent", but the drawing method frozen word-for-word:
Use minimalist two-dimensional black line animation. Show the same consistent stick figure with a hollow circular head, no face, no eyes, no mouth, no hair, no clothing, no filled body, stable proportions, connected line limbs, and uniform medium line weight. Exactly two arms and two legs are visible per figure. Each limb appears only in its current pose: raising or lowering the right arm replaces that arm's previous line, never leaving a resting arm beneath an active arm. The figure stays black; accent colors affect only props and graphic effects.
▲ ① Sound profile, ② character anchor, both marked "reuse verbatim throughout the film".
"When raising the hand, the old arm line is directly replaced, not leaving a resting arm underneath" — this kind of detail is precisely to prevent the extra arm. Writing "same as above" doesn't count; duplicating the whole block three times is most reliable.
Then the storyboard table. Each row is 10 seconds, following "golden hook → break common sense → pull out the inside → truth → elevate and interact", each row broken into three beats, the visual changing every 2–3 seconds, and the character never standing still:
▲ Each segment is 10 seconds, with narrative task, action transition, and narration scheduling all in the table.
At the end of the proposal it stopped and gave three ways forward, even laying out where the subtitles would go. I chose to generate three segments independently — if one segment is unsatisfying, I rerun only that one:
▲ ① Subtitle position; ② "Please confirm, choose one of three"
Step 3: After confirmation, the prompts line up by themselves
I replied:
✅ Proposal approved. Please read the prompt specification and generate 3 independent Omni Flash generation prompts (Phase B), each self-contained, with joining instructions.
It first gave a global continuity block (aspect ratio, style, color scheme, frozen narration, prop-through-line strategy), then output the three prompts in order:
▲ ① Global continuity block; ② CLIP 1 full prompt
Each prompt is self-contained: output spec, canvas definition, character lock, color scheme, composition, first-frame inheritance, three-beat action, verbatim narration, sound profile, BGM, end-frame handoff, negative constraints — 12 items filled in order.
Taking any one segment out by itself, dropping it into any tool that supports prompt-to-video generation, it runs. Not locked to any single vendor.
▲ ① CLIP 3 prompt; ② The right panel has already saved the entire production pack as a document.
The seams between the three segments are also arranged: the first segment's final red square flies out of the right edge, and the second segment opens with the same square falling into the top-left cell to catch it. Not a hard cut; it's action-to-action.
After writing, it quality-checks itself first:
▲ ① Acceptance reminder: focus on checking whether an extra arm has been drawn; ② Subtitle bar
Finally, it saves the entire production pack as a file, 16.9 KB, previewable in the right panel:
▲ The production pack auto-saves; filename and size are listed in the reply.
Note: A mode delivers prompts, not video files. To get an MP4, you have to go further yourself.
Step 4: Render locally to get output
I looked into the official B mode runtime: it requires npm to install canvas and sharp, you bring your own fonts and narration tracks, and you have to run a whole acceptance gate.
The rules are rigorous, but I chose a more direct path — letting WorkBuddy render on this machine:
Final step: render the storyboard into a real video. I've already written the rendering script: D:/.../render_local.py (PIL per-frame drawing + ffmpeg pipe compositing, fully local, no external API calls). Please run this script directly, render a 30-second 16:9 landscape version and a 10-second 9:16 portrait version as two mp4s, then verify the codec and duration with ffprobe and tell me the output file paths.
The script logic is not complex: it reads the storyboard JSON, calculates the character's position and pose every second, draws frame by frame, feeds the frame stream to ffmpeg, and after compositing automatically verifies the codec and reports the duration.
A 30-second piece has over 700 frames; the local CPU whirred for a bit, and two files landed.
It only uses CPU, no video generation model, no TTS.
Its report after completion:
▲ ① Rendering reply: both mp4s pass ffprobe (1280×720 · 30.000s / 720×1280 · 10.000s)
stickman_30s_16x9.mp4 (1280×720, 1.1MB) and stickman_10s_9x16.mp4 (720×1280, 0.43MB), ffprobe verifies H.264 + AAC, duration exact to 30.000 seconds. The first 30-second landscape version is that one.
For the 9:16, I didn't take the landscape and crop a slice. Per the skill's revision rules, changing aspect ratio requires re-layout: title goes into the top safe zone, character and props stack vertically, the pointer changes from a right-side dial to a bottom horizontal bar. The portrait version is a separately rendered one:
One play isn't enough — using the same skill with a different style tier, another piece: Style 2A minimal tech style: a little person in a red hat and yellow T-shirt, a white studio, glowing data cards, 20 seconds explaining "how two pages of meeting notes become three to-dos". The process is identical to the first piece, with the instruction changed to this:
Use A director mode to make a stick-figure video: topic "how two pages of meeting notes become three to-dos", 16:9, 20 seconds, Style 2A minimal tech style (red-hat yellow-shirt little person, white studio, glowing cyan glass UI). Give me the complete director's proposal, stop and wait for my confirmation first.
Switching style tier is not just changing a name. The entire character anchor block changes to the 2A version: red hat, yellow shirt, hollow round head, no realistic facial features, no pom-pom; the color discipline also changes — in Style 1, accent colors only go to props and the character stays black-line; in 2A, the character itself carries color.
Each drawing style is a frozen drawing contract. This piece's output is the second one at the start.
How to choose and my advice
After running the whole set, my judgment is clear.
This solution suits: parents who want to explain reasoning to children; instructors who want to slip animated demonstrations into courseware; science account owners who worry about visuals every day; friends who have installed WorkBuddy but haven't played with skills yet.
It also has boundaries: if you want realistic live-action feel or film-grade camera moves, the stick-figure medium isn't right; if you want full-auto one-click output without confirming a single step, A mode's rhythm will frustrate you; if your computer doesn't even have Python and ffmpeg, you need to add those two first.
Three reminders that can save a lot of time:
Don't treat A mode's prompts as the video
It stops at delivering prompts; to get an MP4 you have to hook up a rendering step yourself. The official B mode environment is not low-barrier; exporting the storyboard and rendering locally saves more trouble.
The character anchor must be frozen verbatim
Writing "the same character" or "same as above" doesn't count; stick figures easily grow an extra hand. Each style tier has its own color discipline; crossing into another color breaks the style.
Switching aspect ratio requires re-layout, not cropping
16:9 is left-center-right scheduling; 9:16 needs vertical stacking; directly cropping the landscape version will push subtitles out of the safe zone. Before uploading images to WeChat official accounts, remember to compress: ASCII filename, leave alt empty, small JPEG under 1200.
What's worth more than speed: every link lands on your own computer. The storyboard is plain-text JSON, the script is local Python, the finished piece is a local ffmpeg-composited MP4, and every file can be taken and modified.
Want to change the topic? Change one sentence. Want to change the drawing style? Pick from twelve style tiers at will. Once this process settles, it becomes your video production line.
From installing the skill to outputting a video: one afternoon. This isn't a concept demo; it's something I just ran in the last two days.
What do you want the stick figure to explain? Leave a comment and I'll pick a few to test in the next issue.
Above is the complete hands-on record; the 6 sets of text can be reused directly. Settle the storyboard JSON and rendering script into a pipeline, and changing a topic is just one sentence — that is your own video production line.