Runway.Cli.Examples

A visual showcase of Runway.Cli + the runway-cli agent skill + Claude Code as the orchestrator.

Each example is a tiny bash wrapper that hands a minimal user-style prompt to claude -p running headless. Claude — guided by the installed runway-cli skill — picks the right Runway CLI commands, generates intermediate assets when needed, and writes a final result.json describing what it made.

No pre-existing assets are required. Every example runs from zero.

---

Showcase

Preview	Example	Prompt	Output	Time	Runway credits
🔊 audio-only (no visual)	audio	"a launch-VO line plus a matching short SFX clip"	1 voice MP3 + 1 SFX MP3 + result.json	~30 s	7 (text-to-speech + sound-effect)
	image	"a minimalist coffee grinder on brushed steel, soft morning light"	1 PNG + result.json	~49 s	7 (1 × runway image)
	product-photoshoot	"a transparent Bluetooth speaker on brushed steel — full social-carousel"	4 publication-ready ad shots + recipe plan + result.json	~3 min	4 (product-photoshoot create --mode social_carousel on gpt_image_2 low-quality)
	mockup	"generate a logo, then place it into 4 advertising mockups via mockup-generator"	1 logo + 4 placement renders + result.json	~3 min	96 (image + mockup-generator workflow)
	haircut	"give someone a virtual haircut: generate a portrait, then restyle with ai-hair-salon"	1 portrait + 4 restyled variations + result.json	~3 min	95 (image + ai-hair-salon workflow)
	virtual-try-on	"generate a person and a leather jacket, then put the jacket on them on a foggy harbor at dawn"	1 person + 1 garment + 4 try-on variations + result.json	~3 min	91 (2 × runway image + virtual-try-on)
	manga	"a 3-page manga about a samurai cat befriending a rival ninja mouse"	6 PNGs + result.json (2 char refs + 4 panels)	~6 min (with the workflow nudge)	163 (1 char image + json-to-manga workflow producing 8 panels)
	storyboard	"a quiet detective scene in a rainy alley"	1 storyboard PNG + result.json	~1 min	21 (1 × storyboard-creator)
	image-to-video	"a ripe peach on marble, soft window light — animate it with a slow push-in"	1 still PNG + 1 animation MP4 + result.json	~3 min	55 (image + gen4-turbo animation)
	short-video	"a short cinematic video of a vintage motorcycle through a foggy mountain pass at sunrise"	3-shot stitched MP4 (18 s) + plan.json + result.json	~6 min	200 (3 × gemini-image3-pro keyframes + 3 × veo3.1-fast shots)
	game-item	"design a jeweled sapphire rune-staff with 4 rarity variants and an animated icon"	4 rarity-tier animated MP4s + result.json	~6 min	675 (game-item-generator renders 4 icons + 4 animated icons on veo3.1)
	wine-label	"design a fictional vineyard 'Stellar Vines' — generate bottle + two labels + hero MP4"	3 source PNGs + composited preview + hero MP4 + result.json	~7 min	1241 (3 × runway image + wine-label-generator with 3 video variations)

Each example's own README has the same layout repeated in detail: Prompt → Inputs → What Claude did → Output → Run it → Cost & runtime.

Cost transparency. The Runway credits column is the measured delta of runway organization get | jq .creditBalance before and after each run — captured by scripts/_runner.sh into meta.json as runway_credits.{before,after,used}. Claude token spend is captured separately in each run's meta.json (claude_cost_usd) and noted in the per-example README, but kept out of this table to keep it scannable.

---

Setup

Prereqs: Claude Code, .NET 10+, Node.js (for npx), jq, a Runway API key.

git clone https://github.com/tryAGI/Runway.Cli.Examples.git
cd Runway.Cli.Examples
cp .env.example .env
# Edit .env and set RUNWAY_API_KEY
./scripts/setup.sh

How the skill is installed

scripts/setup.sh runs the install command documented in the upstream Runway repo (Runway/README.md#use-as-an-agent-skill):

npx skills add tryAGI/Runway -a claude-code -y

This drops the skill at .claude/skills/runway-cli/SKILL.md (provided by the skills.sh ecosystem). Headless claude -p invocations in this repo pick it up automatically. The skill file itself is not committed — it is re-installed by setup.sh on a fresh clone. The auto-generated skills-lock.json records the exact source + hash for reproducibility.

Run an example

./examples/audio/run.sh              # ~30 s, $0.15 Claude,    7 Runway credits
./examples/product-photoshoot/run.sh # ~3 min,$0.19 Claude,    4 Runway credits (gpt_image_2 low-q)
./examples/image/run.sh              # ~49 s, $0.15 Claude,    7 Runway credits
./examples/storyboard/run.sh         # ~1 min,$0.16 Claude,   21 Runway credits
./examples/image-to-video/run.sh     # ~3 min,$0.20 Claude,   55 Runway credits
./examples/haircut/run.sh            # ~3 min,$0.31 Claude,   95 Runway credits
./examples/virtual-try-on/run.sh     # ~3 min,$0.66 Claude,   91 Runway credits
./examples/mockup/run.sh             # ~3 min,$0.20 Claude,   96 Runway credits
./examples/manga/run.sh              # ~6 min,$0.58 Claude,  163 Runway credits (workflow-nudged)
./examples/short-video/run.sh        # ~6 min,$0.61 Claude,  200 Runway credits
./examples/game-item/run.sh          # ~6 min,$0.32 Claude,  675 Runway credits (4 animated MP4s)
./examples/wine-label/run.sh         # ~7 min,$0.65 Claude, 1241 Runway credits (video!)

Output lands in output/<example>/<ISO-timestamp>/:

output/<example>/<timestamp>/
├── result.json     # the final JSON Claude produced
├── transcript.json # the raw `claude -p --output-format json` envelope
├── meta.json       # claude version, runway version, cost, session id
└── assets/         # generated PNGs (and MP4s for video examples)

Each example also commits a sample-output/ directory containing real artifacts from a real run, so the README previews are honest evidence rather than mockups.

Examples index

Workflow	Status	Path
audio (text-to-speech + sound-effect)	shipped	examples/audio/
image (json-to-image)	shipped	examples/image/
product-photoshoot	shipped	examples/product-photoshoot/
storyboard (storyboard-creator)	shipped	examples/storyboard/
image-to-video	shipped	examples/image-to-video/
haircut (ai-hair-salon)	shipped	examples/haircut/
virtual-try-on	shipped	examples/virtual-try-on/
mockup (mockup-generator)	shipped	examples/mockup/
manga (json-to-manga)	shipped	examples/manga/
short-video	shipped	examples/short-video/
game-item (game-item-generator)	shipped	examples/game-item/
wine-label (wine-label-generator)	shipped	examples/wine-label/
video (text-to-video direct)	planned	examples/video/
marketplace-cards	planned	examples/marketplace-cards/
ad-video	planned	examples/ad-video/
avatar	planned	examples/avatar/
soul-id	planned	examples/soul-id/
character-creator	planned	examples/character-creator/
b-roll-generator	planned	examples/b-roll-generator/
backplate-generator	planned	examples/backplate-generator/
human-pose-replication	planned	examples/human-pose-replication/
(more named workflows)	planned	—

The goal is one example per Runway CLI workflow — each with the same showcase layout so you can scan the whole repo and see Runway's surface area at a glance.

Reproducibility

Runs are non-deterministic. Each run captures a meta.json with:

• claude CLI version
• runway (Runway.Cli) version
• total cost in USD
• Anthropic session id (for replay/debugging)

Contributing a new example

Every example follows the same six-section README so the repo reads as a uniform showcase:

1. The prompt — one or two sentences in user voice, telling Claude what, not how
2. Inputs — env vars, no pre-existing assets
3. What Claude did — the workflow Claude orchestrated
4. Output — embedded PNGs/MP4s + a result.json snippet (real artifacts in sample-output/)
5. Run it — the exact command
6. Cost & runtime — observed numbers from a real run

To add one:

1. cp -r examples/image examples/<workflow-name>/
2. Rewrite prompt.md for the new workflow
3. Run ./examples/<workflow-name>/run.sh once and commit a curated subset of the output into sample-output/
4. Update the README following the six-section structure
5. Add a row to the Examples index table above

The skill teaches Claude how to drive Runway. Your example just needs to ask, in plain language, and capture the result.

Related

• tryAGI/Runway — the SDK, CLI, and skill
• skills.sh — agent-skill ecosystem this repo consumes
• Claude Code — the agent runtime