m18

Bikin visual, animasi, diagram, slide, dan video retro dari teks. Semua tool di sini opsional — load cuma pas operator minta konten kreatif. Bukan jalur dana, gak ada gerbang governor. Output = fi

SKILL.md

GitHub

m18 — Creative & Media Generation (NEW in v4.1)

Bikin visual, animasi, diagram, slide, dan video retro dari teks. Semua tool di sini opsional — load cuma pas operator minta konten kreatif. Bukan jalur dana, gak ada gerbang governor. Output = file artefak (komplemen m8).

Capability Engine Output
Image / workflow gen ComfyUI (API) PNG/batch
Math animation Manim Community MP4/GIF
Diagram / sketch Excalidraw (.excalidraw JSON) SVG/PNG
ASCII / retro video chafa / ascii-video terminal/MP4
Web design system desain-token + komponen HTML/CSS/Tailwind
Slide polish Felo-style content→deck PPTX/HTML

1. ComfyUI advanced workflow (API mode)

ComfyUI headless = node-graph JSON di-POST ke /prompt. Jangan klik UI; drive via API biar reproducible & batch-able.

# pip install httpx
import httpx, json, time, uuid

COMFY = "http://127.0.0.1:8188"

def run_workflow(graph: dict, timeout=300) -> list[str]:
    """POST graph JSON (export 'Save (API Format)' dari ComfyUI), poll history, return file paths."""
    cid = uuid.uuid4().hex
    r = httpx.post(f"{COMFY}/prompt", json={"prompt": graph, "client_id": cid}, timeout=30)
    r.raise_for_status()
    pid = r.json()["prompt_id"]
    t0 = time.time()
    while time.time() - t0 < timeout:
        h = httpx.get(f"{COMFY}/history/{pid}", timeout=15).json()
        if pid in h:
            outs = []
            for node in h[pid]["outputs"].values():
                for img in node.get("images", []):
                    outs.append(f"{COMFY}/view?filename={img['filename']}&subfolder={img.get('subfolder','')}&type={img['type']}")
            return outs
        time.sleep(1.5)
    raise TimeoutError(f"workflow {pid} > {timeout}s")

Pola: export graph dari ComfyUI via Save (API Format), simpan sebagai template, lalu patch node yang perlu (prompt text, seed, model) sebelum POST. Batch = loop seed via m12 (p-limit/concurrency). Heavy GPU → jalankan ComfyUI di VPS GPU, drive dari agent via SSH tunnel.

→ Reference workflow templates & node patching helper: tools/scene_prep.py (juga dipakai m19 buat 3D scene).

2. Manim — animasi matematika/penjelasan

# pip install manim   (butuh ffmpeg + LaTeX untuk teks math)
from manim import *
class Explain(Scene):
    def construct(self):
        eq = MathTex(r"e^{i\pi} + 1 = 0")
        self.play(Write(eq)); self.wait(1)
        self.play(eq.animate.to_edge(UP))
        self.play(Create(Axes(x_range=[-3,3], y_range=[-2,2])))

Render: manim -qm scene.py Explain (-qm medium, -qh high, -ql cepat draft). Output media/videos/.../Explain.mp4. Buat konsep teknis → animasi step-by-step lebih nempel dari teks. Pasangkan ke m3 (naskah voiceover).

3. Excalidraw — diagram hand-drawn

Excalidraw = JSON {type:"excalidraw", elements:[...]}. Generate programmatically untuk arsitektur/flow, bukan gambar manual:

def excalidraw_box(x, y, w, h, text, eid):
    rect = {"type":"rectangle","id":eid,"x":x,"y":y,"width":w,"height":h,
            "strokeColor":"#1e1e1e","backgroundColor":"#a5d8ff","roughness":1,"seed":1}
    label = {"type":"text","id":eid+"_t","x":x+12,"y":y+h/2-10,"text":text,
             "fontSize":16,"containerId":eid,"seed":2}
    return [rect, label]
# rangkai boxes + arrows → {"type":"excalidraw","version":2,"elements":[...]}

Simpan .excalidraw → buka di excalidraw.com / VS Code plugin → export SVG/PNG. Bagus buat diagram arsitektur sistem (mis. visualisasi router AGENTS.md).

4. ASCII / retro video

# image → ASCII berwarna
chafa input.png --size 80x40
# video → ASCII playback (terminal)
ffmpeg -i clip.mp4 -vf fps=12,scale=120:-1 frames/%04d.png && for f in frames/*.png; do chafa "$f" --size 120x50; sleep 0.08; clear; done
# render balik ke MP4 retro (CRT vibe): ffmpeg + curves/noise filter
ffmpeg -i clip.mp4 -vf "curves=preset=vintage,noise=alls=8:allf=t" retro.mp4

5. Web design systems (Stripe / Linear / Vercel vibe)

Jangan reinvent — pakai design token yang sudah teruji. Cluster ke m9 buat implementasi React/Tailwind.

Sistem Sinyal visual Token inti
Stripe gradient halus, depth, trust indigo #635bff, soft shadow, 8px grid
Linear dark-first, crisp, fast #5e6ad2, mono accents, 1px borders
Vercel mono B/W, geometric pure black/white, Geist font, harsh contrast

Output design token JSON → feed ke m9 untuk komponen.

Aesthetic judgment & design approval

Sebelum bilang "udah bagus", lewatin checklist — jangan ACC buta:

✅ Hierarki jelas? (mata tau harus liat apa duluan)
✅ Spacing konsisten? (grid 4/8px, bukan angka random)
✅ Kontras cukup? (teks ke-baca, WCAG AA)
✅ ≤2 typeface, ≤1 warna aksen dominan? (restraint > ramai)
✅ Alignment rapi? (gak ada elemen "ngambang")
✅ Lulus squint-test? (dipicingin mata, struktur tetap kebaca)

Ada yang gagal → tunjuk spesifik + perbaikan, jangan "kurang menarik" doang. Approval final yang ireversibel (publish/print) → sign-off operator (x7).

6. Felo-style: content → polished deck

Pipeline: outline (LLM via model_registry) → struktur slide → render. Untuk PPTX gunakan skill pptx bila tersedia, atau HTML reveal.js untuk web deck.

brief → m3 (outline + hook) → m18 (layout + design token) → m8/pptx (export)

Satu kalimat brief → deck rapi. Jangan kasih bullet mentah; setiap slide = 1 ide + 1 visual.

7. Canvas design & visual art workflows

Komposisi visual programmatic (poster, social card, thumbnail, generative art) tanpa buka editor:

# pip install pillow — komposit cepat: bg + teks + layer
from PIL import Image, ImageDraw, ImageFont
img = Image.new("RGB", (1200, 630), "#0A2540")            # OG card 1200×630
d = ImageDraw.Draw(img)
d.text((60, 260), "SUPERAGENT", fill="#635BFF", font=ImageFont.truetype("Inter-Bold.ttf", 72))
img.save("card.png")
  • Generative/AI art → routing ke ComfyUI (§1) buat raster, atau SVG programmatic buat vektor.
  • Layout berbasis grid + design token (§5) biar konsisten, bukan tebak posisi.
  • Infografis/carousel: 1 ide per panel, hierarki angka→label, ikon konsisten. Slide deck → §6 (Felo).
  • Aesthetic judgment (checklist di atas) sebelum ACC. Batch (banyak card) → m12.
  • Image prompt engineer: buat prompt Midjourney/Flux/Canva (subjek + gaya + komposisi + lighting + negative) → render via ComfyUI (§1). Repurpose 1 konten → banyak visual platform = pipeline m29.

Catatan token & ops

  • Semua engine di sini lokal/self-host atau keyless — gak nambah always-on cost.
  • ComfyUI/Manim butuh GPU/ffmpeg/LaTeX → deploy via m2, drive jarak jauh.
  • Konten kreatif bukan operasi berisiko → gak lewat governor, gak butuh R9 gate.
  • Combo umum: m18 + m3 (naskah) + m8 (export file) + m9 (web render).