Riff

Plan mode: If you are planning work, this entire skill is ONE plan step: "Invoke /vibes:riff". Do not decompose the steps below into separate plan tasks.

Display this ASCII art immediately when starting:

░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░▒▓███████▓▒░░▒▓████████▓▒░░▒▓███████▓▒░
░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░      ░▒▓█▓▒░
 ░▒▓█▓▒▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░      ░▒▓█▓▒░
 ░▒▓█▓▒▒▓█▓▒░░▒▓█▓▒░▒▓███████▓▒░░▒▓██████▓▒░  ░▒▓██████▓▒░
  ░▒▓█▓▓█▓▒░ ░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░             ░▒▓█▓▒░
  ░▒▓█▓▓█▓▒░ ░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░             ░▒▓█▓▒░
   ░▒▓██▓▒░  ░▒▓█▓▒░▒▓███████▓▒░░▒▓████████▓▒░▒▓███████▓▒░

Vibes Riff Generator

Generate multiple app variations in parallel. Each riff is a different INTERPRETATION - different ideas, not just styling.

Terminal or Editor UI?

Detect whether you're running in a terminal (Claude Code CLI) or an editor (Cursor, Windsurf, VS Code with Copilot). Terminal agents use AskUserQuestion for all input. Editor agents present requirements as a checklist comment, wait for user edits, then proceed. See the vibes skill for the full detection and interaction pattern.

Workflow

Step 1: Gather ALL Requirements Upfront

Use AskUserQuestion to collect all config at once.

Question 1: "What kind of app do you want to explore? (broad/loose is fine)"
Header: "Theme"
Options: User enters via "Other"

Question 2: "Describe the visual style - colors, mood, aesthetic"
Header: "Visual"
Options: ["Warm sunset tones", "Clean minimal white", "Neon cyberpunk", "Other (describe)"]

Question 3: "How many variations should I generate?"
Header: "Count"
Options: ["3 (recommended)", "5", "7", "10"]

After receiving all answers, proceed immediately to Step 2 - no more questions.

Note: If the frontend-design skill is available, use it for enhanced visual design quality.

Step 2: Create Riff Directories

mkdir -p riff-1 riff-2 riff-3 ...

Step 3: Generate Riffs in Parallel

Use the bundled script to generate riffs in parallel. Each script instance calls claude -p (uses subscription tokens) and writes directly to disk.

Generate riffs in parallel based on user's count:

VIBES_ROOT="${CLAUDE_PLUGIN_ROOT:-$(dirname "$(dirname "${CLAUDE_SKILL_DIR}")")}"
# For each N from 1 to ${count}:
bun "$VIBES_ROOT/scripts/generate-riff.js" "${prompt}" N riff-N/app.jsx "${visual}" &

# Then wait for all
wait
echo "All ${count} riffs generated!"

Example for count=3:

VIBES_ROOT="${CLAUDE_PLUGIN_ROOT:-$(dirname "$(dirname "${CLAUDE_SKILL_DIR}")")}"
bun "$VIBES_ROOT/scripts/generate-riff.js" "productivity apps" 1 riff-1/app.jsx "warm sunset oranges and soft creams" &
bun "$VIBES_ROOT/scripts/generate-riff.js" "productivity apps" 2 riff-2/app.jsx "warm sunset oranges and soft creams" &
bun "$VIBES_ROOT/scripts/generate-riff.js" "productivity apps" 3 riff-3/app.jsx "warm sunset oranges and soft creams" &
wait

Why this works:

• Each script calls claude -p "..." → uses subscription tokens
• Script writes directly to disk → no tokens flow through main agent
• Background processes (&) run in parallel → true concurrency
• Main agent only sees "✓ riff-N/app.jsx" output → minimal tokens

Step 4: Assemble HTML

Convert each app.jsx to a complete index.html:

VIBES_ROOT="${CLAUDE_PLUGIN_ROOT:-$(dirname "$(dirname "${CLAUDE_SKILL_DIR}")")}"
bun "$VIBES_ROOT/scripts/assemble-all.js" riff-1 riff-2 riff-3 ...

Step 5: Evaluate & Rank

Read the pitch.md files (NOT the full code) for fast evaluation:

# Read pitch files - contains reasoning about theme, colors, functionality
cat riff-*/pitch.md

# Also read BUSINESS comments for app names (just first 10 lines of each)
head -10 riff-*/app.jsx

pitch.md contains the model's reasoning about theme, colors, and design choices. BUSINESS comment (top of app.jsx) contains: name, pitch, customer, revenue.

Then create RANKINGS.md using this format:

# Riff Rankings: [Theme]

## Summary

| Rank | App Name | Score | Best For |
|------|----------|-------|----------|
| 1 | [Name] | XX/50 | [one-liner] |
| 2 | [Name] | XX/50 | [one-liner] |
| ... | ... | ... | ... |

## Detailed Scores

### #1: [App Name] (riff-N)
| Criterion | Score | Notes |
|-----------|-------|-------|
| Originality | X/10 | [Why unique or derivative] |
| Market Potential | X/10 | [Target audience size, demand] |
| Feasibility | X/10 | [Technical complexity, time to build] |
| Monetization | X/10 | [Revenue model viability] |
| Wow Factor | X/10 | [First impression, engagement] |
| **Total** | **XX/50** | |

[Repeat for each riff]

## Recommendations

- **Best for solo founder:** [riff-N] - [reason]
- **Fastest to ship:** [riff-N] - [reason]
- **Most innovative:** [riff-N] - [reason]
- **Best monetization:** [riff-N] - [reason]

Scoring Guidelines:

• 1-3: Poor - significant issues
• 4-5: Below average - notable weaknesses
• 6-7: Average - meets expectations
• 8-9: Good - stands out positively
• 10: Excellent - exceptional in this criterion

Step 6: Generate Gallery

Create index.html gallery page with:

• Dark theme (#0a0a0f background)
• Glass-morphism cards with purple/cyan accents
• Each card: rank badge, app name, pitch, score bar, "Launch →" link
• Responsive grid layout
• Self-contained with inline styles

Step 7: Present Results

Generated ${count} riffs for "${prompt}":
#1: riff-X - App Name (XX/50)
#2: riff-Y - App Name (XX/50)
...

Open index.html for gallery, or browse riff-1/, riff-2/, etc.

Plugin Directory

Scripts resolve the plugin root via VIBES_ROOT, which tries ${CLAUDE_PLUGIN_ROOT} first (set by Claude Code) and falls back to deriving the path from ${CLAUDE_SKILL_DIR} (two directories up). Set it at the top of each bash block: VIBES_ROOT="${CLAUDE_PLUGIN_ROOT:-$(dirname "$(dirname "${CLAUDE_SKILL_DIR}")")}"

What's Next?

After presenting rankings, guide the user with AskUserQuestion:

Question: "You have ${count} variations ranked. What would you like to do?"
Header: "Next"
Options:
- Label: "Develop the #1 ranked app"
  Description: "Take the top-ranked variation and continue building it. I'll copy riff-1/app.jsx to your working directory so you can iterate with the full vibes feature set - add functionality, refine the design, make it yours."

- Label: "Let me pick a different one"
  Description: "The rankings are just my analysis - you might see something special in another variation. Tell me which riff number speaks to you and I'll set it up for development."

- Label: "Turn winner into SaaS (/sell)"
  Description: "Ready to monetize the best one? Sell adds authentication via Pocket ID with passkeys, subscription billing, and multi-tenant database isolation. Each customer gets their own subdomain with their own data."

- Label: "Deploy a riff as demo (/cloudflare)"
  Description: "Want to share one quickly? Deploy puts any variation live on Cloudflare Workers within minutes. Great for getting feedback before committing to further development."

- Label: "I'm done for now"
  Description: "Wrap up this session. All riffs are saved in riff-1/, riff-2/, etc. - browse them locally or come back later to continue."

After user responds:

• "Develop #1" → Copy riff-1/app.jsx to ./app.jsx, confirm ready for vibes iteration
• "Pick different" → Ask which riff number, then copy that one
• "Turn into SaaS" → Copy riff-1/app.jsx, auto-invoke /vibes:factory skill
• "Deploy demo" → Ask which riff, auto-invoke /vibes:cloudflare skill with that riff's index.html
• "I'm done" → Confirm riffs saved, provide directory listing