puntorigen/skills

Local-first agent skills for Cursor and other AI coding agents. Almost every skill runs on your machine — no cloud APIs, no API keys. Most are optimized for Apple Silicon Macs (MLX / MPS / Metal). Even brand-logo-kit, the one skill that can reach the cloud, is local-first: it prefers on-device FLUX.2 Klein for logos and only falls back to a Gemini/OpenRouter key (which it reuses from your environment, never bundling one) when local can't run.

Install with the skills CLI:

npx skills add puntorigen/skills              # install all
npx skills add puntorigen/skills@talking-head # one skill
npx skills add puntorigen/skills -g -y        # global install, skip prompts

Browse: skills.sh/puntorigen/skills

Available skills

image-gen

Generate hyper-realistic images from a text prompt at any resolution, locally with mflux (Z-Image-Turbo or FLUX.2-klein). Optional SeedVR2 upscaling.

npx skills add puntorigen/skills@image-gen -g -y

Requires: Apple Silicon Mac, uv, ~6 GB disk for default model.

---

voice-clone-narration

Clone a voice from a short sample or design one from a description. Generate expressive MP3 narrations in English or Spanish with Chatterbox / Qwen3-TTS.

npx skills add puntorigen/skills@voice-clone-narration -g -y

Requires: Python 3.11, ffmpeg, ~5 GB disk. Voice design is Apple Silicon only.

---

bg-music

Generate royalty-free instrumental background music from a text brief with ACE-Step 1.5. Mix under a voiceover with automatic ducking.

npx skills add puntorigen/skills@bg-music -g -y

Requires: uv, git, ffmpeg, ~10 GB disk. Best on Apple Silicon.

---

sound-effects

Generate sound effects, foley, and ambience (door slams, footsteps, rain, wind, UI blips, whooshes) from a text prompt with Stable Audio Open Small (via mlx-audiogen). Tuned for SFX and field recordings, up to ~11s stereo.

npx skills add puntorigen/skills@sound-effects -g -y

Requires: Apple Silicon Mac, uv, ffmpeg, ~1-2 GB disk. Weights are public (no HF account/token needed); under the Stability AI Community License (free under $1M revenue).

---

audio-theater

Turn a dialogue, story, or idea into finished multi-character audio — a radio drama, a two-host podcast, or clean lip-sync clips — by orchestrating voice-clone-narration (one voice per character), sound-effects (foley), and bg-music (score), then mixing with ffmpeg (ducking + optional stereo spatialization). Delivers the full mix plus music/no-music stems and a timecoded transcript.

npx skills add puntorigen/skills@audio-theater -g -y

Requires: ffmpeg + the voice-clone-narration skill (required); optional sound-effects and bg-music for SFX/music cues. No venv of its own.

---

talking-head

Turn a portrait image + narration audio into a lip-synced talking-head MP4 with JoyVASA and LivePortrait. Composes with image-gen and voice-clone-narration.

npx skills add puntorigen/skills@talking-head -g -y

Requires: Apple Silicon Mac, uv, git, ffmpeg, ~6 GB disk.

---

video-to-splat

Convert an mp4 walkthrough into a 3D Gaussian splat (PLY/SOG) and preview it in a bundled Aholo viewer. Full local photogrammetry pipeline.

npx skills add puntorigen/skills@video-to-splat -g -y

Requires: Apple Silicon Mac (macOS 14+), uv, ffmpeg, node, Chrome/Edge 134+.

---

object-to-3d

Turn an mp4 of an orbited object into a clean, browser-navigable Gaussian splat and a watertight, millimeter-scaled STL/GLB mesh for 3D printing. Extends the video-to-splat pipeline with automatic splat cleanup and printable mesh extraction - Poisson reconstruction plus base-down orientation, voxel solidify and a flat print base, with in-browser Splat and Print previews (open3d + trimesh + scikit-image).

npx skills add puntorigen/skills@object-to-3d -g -y

Requires: Apple Silicon Mac (macOS 14+), uv, ffmpeg, node, a WebGL2 browser.

---

generate-web-skills

Record a real Chrome session (HAR + UI steps), distill it into a reusable lesson, then replay the action with different parameters or record UI proof.

npx skills add puntorigen/skills@generate-web-skills -g -y

Requires: Node.js, Google Chrome, python3, ffmpeg (for mp4 proof). See generate-web-skills/README.md.

---

nextjs-onboarding-tours

Add a complete, task-oriented onboarding layer to any Next.js app or admin panel — a first-login welcome wizard, re-launchable guided tours per area (driver.js), dashboard quick actions, an in-app help center, and a floating help button, with per-user persistence. Everything auto-shows at most once and stays one click away.

npx skills add puntorigen/skills@nextjs-onboarding-tours -g -y

Requires: a Next.js app (App Router; great with Payload CMS admin, NextAuth, or a custom auth context). Adds driver.js (~5 kB) via pnpm — no models, no cloud, no API keys.

---

edit-docx

Inspect and edit Microsoft Word .docx files while preserving formatting, styles, and layout — run-aware find/replace, insert paragraphs, edit table cells, add rows, delete elements — all locally with python-docx.

npx skills add puntorigen/skills@edit-docx -g -y

Requires: Python 3.9+ (any OS). Optional uv for a faster install.

---

pdf-documents

Read PDFs with real layout understanding (Docling → structured JSON/Markdown, tables, sections, OCR) and create professional multi-page PDFs from a JSON config (fpdf2 → titles, styled tables, images, lists, headers/footers) — all locally.

npx skills add puntorigen/skills@pdf-documents -g -y

Requires: Python 3.9+ (any OS), ~1-2 GB disk for Docling models (first read).

---

j-space

Build and use an external, persistent mental workspace for an agent, modeled on the "J-space" from Anthropic's global-workspace research. Turn a topic (Perplexity/web research) or a codebase (scan) into a weighted concept graph, compile it into a matrix with local embeddings, then load, query (spreading activation), edit, checkpoint, and even hypnotize the workspace — pin concepts, install triggers, and write an always-on induction rule that primes the agent every session.

npx skills add puntorigen/skills@j-space -g -y

Requires: Python 3.11+, uv, ~500 MB disk for the venv + all-MiniLM-L6-v2 embedding model (public, no HF account; MPS or CPU, any platform). See j-space/README.md.

---

brand-logo-kit

Generate a logo and a matching, brand-consistent asset set — wordmarks, monograms, app icons, favicons, social avatars, banners, seamless patterns, and spot illustrations. Pick a logo, set the name in a real font (wordmark.py), extract the palette, and keep the whole set cohesive with a shared --look.

npx skills add puntorigen/skills@brand-logo-kit -g -y

Local-first, with a cloud fallback. By default it renders on-device via the local image-gen skill (FLUX.2 Klein, Apple Silicon) — no key, no cloud — chosen automatically when there's an Apple Silicon Mac with enough disk for the weights. When local can't run it falls back to the cloud (Gemini / OpenRouter Nano Banana Pro), bundling no API key: it auto-discovers GEMINI_API_KEY / OPENROUTER_API_KEY from your environment or another installed skill and caches it outside the repo. --look presets steer the finish, and wordmark.py sets brand text from genuine fonts (no misspelled diffusion text).

Requires: Python 3.9+ (any OS). For the preferred local path: an Apple Silicon Mac + the image-gen skill. For the cloud fallback: a Gemini or OpenRouter API key (auto-discovered). Optional uv for a faster install.

End-to-end reel workflow

These skills compose into a fully local content pipeline:

image-gen  →  avatar.png
voice-clone-narration  →  narration.mp3
talking-head  →  talking-head.mp4
bg-music  →  reel-audio.mp3 (optional mix)

# or a full audio drama / podcast:
audio-theater  →  orchestrates voice-clone-narration + sound-effects + bg-music
               →  final.mp3 (+ music / no-music stems) + transcript.md
               →  clean per-line clips → talking-head (on-camera lip-sync)

Install the set:

npx skills add puntorigen/skills \
  -s image-gen,voice-clone-narration,bg-music,sound-effects,audio-theater,talking-head \
  -g -y

Each skill stores models and outputs outside the repo (under ~/.* home dirs). First run of scripts/setup_env.sh downloads weights and may take several minutes.

Example: a 3-voice spatial radio drama

examples/audio-theater-storm/ is a complete, reproducible audio-theater run — a ~50-second storm scene with three distinct designed voices (no reference clips), acted with per-line emotion, and four spatialized sound effects, mixed to a stereo stage. 100% local, no cloud, no Hugging Face account.

▶ Hear it — download / play the 51-second demo (MP4):

audio-theater-storm-demo.mp4

Static preview of the mix — left channel (blue) over right (orange). They differ because each voice and one-shot sits at its own pan/distance; e.g. the left-heavy burst is the door slam on Mara's side of the stage.

The story and cues are authored as small JSON files; the agent then drives the sibling skills:

OUT=examples/audio-theater-storm
SC=audio-theater/scripts

python3 $SC/setup_cast.py       --out $OUT                             # design 1 voice / character
python3 $SC/generate_voices.py  --script $OUT/script.json --out $OUT --seed 700  # -> dialogue.wav + lines.json
python3 $SC/generate_sfx.py     --cues   $OUT/cues.json   --out $OUT   # ambient bed + one-shots
python3 $SC/mix_spatial.py      --out $OUT                             # -> final.mp3 (stereo stage)
python3 $SC/build_transcript.py --out $OUT                             # -> transcript.md

Each line is acted: an emotion (+ intensity) drives Chatterbox's exaggeration/cadence, and a theater expressiveness boost widens the dynamic range so calm lines sit calm and the panic actually spikes:

// script.json — cast seated on the stage, each line given a delivery
"expressiveness": 1.25,
"characters": [
  { "name": "Narrator", "role": "narration", "tone": "warm", "stage": { "pan": 0.0,  "distance": 0.12 } },
  { "name": "Mara",     "persona": "weathered lighthouse keeper, low and steady", "tone": "firm",
    "stage": { "pan": -0.4, "distance": 0.2 } },
  { "name": "Tomas",    "persona": "nervous young apprentice, higher and breathy", "tone": "nervous",
    "stage": { "pan": 0.45, "distance": 0.22 } }
],
"lines": [
  { "index": 3, "speaker": "Tomas", "text": "I'm trying! The latch won't hold!", "emotion": "panicked", "intensity": 0.8 }
]

// cues.json — a hard-left door slam on Mara's side (short, concrete prompt + quality knobs)
{ "id": "door", "type": "oneshot", "description": "heavy wooden door slams shut, sharp wooden bang",
  "start": 17.9, "gen_seconds": 3, "steps": 24, "sampler": "rk4",
  "negative_prompt": "music, animal, roar, voice", "spatial": { "pan": -0.55, "distance": 0.15 } }

The rendered transcript.md and the input JSON are committed; the audio (*.mp3/*.wav) is git-ignored — run the commands above to hear it. See the example README for the full walkthrough.

Repo layout

skills/
├── README.md
├── skills.sh.json          # optional grouping for skills.sh
├── image-gen/
│   ├── SKILL.md            # required — agent instructions + frontmatter
│   ├── REFERENCE.md        # optional deep reference
│   └── scripts/
├── voice-clone-narration/
├── bg-music/
├── sound-effects/
├── audio-theater/
├── talking-head/
├── video-to-splat/
├── object-to-3d/
├── generate-web-skills/
├── nextjs-onboarding-tours/
├── edit-docx/
├── pdf-documents/
├── j-space/
├── brand-logo-kit/         # local-first (FLUX.2 Klein) + cloud fallback — logos & brand assets
└── examples/
    └── audio-theater-storm/   # worked audio-theater example (script + cues + transcript)

Verify locally before publishing

# From this repo root:
npx skills add . --list

# Dry-run install one skill:
npx skills add .@talking-head -g -y

Development

Scaffold a new skill:

npx skills init my-new-skill

Each SKILL.md needs YAML frontmatter with name and description. See agentskills.io and existing skills here for conventions.

License

Skill instructions and scripts in this repo are MIT unless noted otherwise. Bundled model weights retain their upstream licenses (ACE-Step MIT, Chatterbox MIT, Qwen3-TTS Apache-2.0, Z-Image/FLUX Apache-2.0, etc.) — see each skill's REFERENCE.md.

Note on sound-effects: Stable Audio Open Small weights are under the Stability AI Community License — free for research and for individuals/organizations under $1M USD annual revenue, but not as permissive as the Apache/MIT models above. The audio-theater skill inherits this when it generates SFX. Review before commercial use.