Riff: Sustained Divergent Exploration

Overview

Explore an idea space without convergence pressure. Where ideation expands possibilities toward a vision and refinement contracts them to a core, riffing stays in divergent space — branching, combining, inverting, and surprising. The output is a branching outline of everything explored, not a decision about what to build.

Riff is standalone. It can be used before ideation (to generate raw material), after refinement (to re-expand a stripped idea), or completely independently. A riff session's output can naturally feed into other skills, but transitions are always the user's choice, never automatic. Riff never suggests "now let's move to ideation" or "this is ready for refinement." It stays in its lane: divergence.

<HARD-GATE> The gate is on shifting from ideation to engineering, NOT on idea content. Ideas expressed through implementation details are welcome — app concepts naturally include technical specifics (Bluetooth proximity detection, push notifications, APIs, webhooks, ML classifiers) because the idea and its mechanism are often inseparable.

What IS prohibited: designing systems, writing code, proposing architecture, creating implementation plans, suggesting tech stacks, scaffolding projects, or invoking implementation skills.

The test: "Is Claude generating new ideas, or has it stopped riffing to start building?" If the answer is building, stop. </HARD-GATE>

When to Use

• User wants to explore a theme, concept, or idea space without committing to a direction
• User wants creative divergence — branches, angles, unexpected connections
• User has a seed (word, phrase, paragraph, or document) and wants to see where it could go
• User says "riff on...", "what could we do with...", "let's brainstorm about..."

When NOT to Use

• User wants to develop a specific idea into a vision document → use ideas:ideation
• User wants to strip an idea to its core or pressure-test it → use ideas:refine
• User is ready to design or implement something → use the superpowers:brainstorming skill for design, or superpowers:writing-plans for implementation planning
• User wants a structured requirements gathering process

Input Handling

Bare theme (word or phrase)

Acknowledge theme, immediately throw out 2-3 opening angles. No clarifying questions. "Neighborhood apps" → start riffing on hyperlocal social, ambient awareness, micro-commerce, whatever comes first. Understanding deepens through riffing, not before it.

Seed + context (theme plus starting thoughts)

Read context, riff into the gaps — angles they haven't mentioned. If the user says "I've been thinking about neighborhood apps for elderly residents," don't repeat their framing. Go where they haven't: intergenerational exchange, passive safety nets, oral history capture, skill-sharing economies.

Existing artifact (vision doc, blog post, README)

Read document, riff outward — what's adjacent to the vision, what's assumed but unexamined, what happens if you invert a key premise. A vision doc about a todo app? Riff on: what if completion wasn't the goal, what if tasks were social, what if the app fought against productivity.

Key principle: No intake phase. Start riffing immediately. Understanding deepens through riffing, not before it.

Technique Toolkit

Four groups of techniques, deployed based on conversational energy and need. Full descriptions in the reference files — summaries here for quick selection.

Reframing (7) — change the lens

See references/techniques-reframing.md for full descriptions.

Technique	What it does	When to deploy
Perspective shift	View through different lens — first-time user, skeptic, competitor, different industry/culture. Includes designing FOR a different audience.	Conversation settled into one viewpoint
Word pivot	Pick key word, swap with synonyms/metaphors. "Network → ecosystem → organism → market"	Framing feels fixed
Metaphor mutation	Evolve a metaphor through successive steps. "Marketplace → auction → game show"	Analogy productive but could go further
Medium shift	Change form entirely. "What if this was a game? A physical object? A protocol?"	Idea trapped in its current form
Emotion-driven ideation	Start from feeling, not function. "What would a delightful version look like?"	Ideation too function-first
Narrative seeding	Turn into a story. "A user wakes up and..."	Abstract thinking isn't producing
Misuse exploration	"What's the weirdest way someone could use this?"	Intended use feels too narrow

Mutation (8) — change the idea

See references/techniques-mutation.md for full descriptions.

Technique	What it does	When to deploy
Inversion	"What if the opposite were true?" Flip assumptions.	Conversation too agreeable
Subtraction	Remove the most obvious element. What's left?	Idea feels bloated or conventional
Exaggeration	Take to 10x/100x scale — but keep plausible.	Ideas safe and incremental
Signal amplification	Pick weak signal, make it the center. "This tiny feature becomes the whole product."	Small detail deserves more attention
Remove the goal	Drop original objective. "If we didn't care about X, what would we build?"	Goal might be wrong or too narrow
Constraint game	Add arbitrary constraint. "What if it had to work offline?"	Exploration broad but shallow
Constraint explosion	Add many conflicting constraints simultaneously.	Single constraints produce predictable results
Future-back ideation	Start from solved future. "In 2035 this is solved — how?"	Current constraints blocking imagination

Connection (4) — bring something in

See references/techniques-connection.md for full descriptions.

Technique	What it does	When to deploy
Analogy	Draw parallels from unrelated domains (biology, music, urban planning, cooking).	Idea stuck in its own domain
Combination	Smash two session ideas together. Includes grafting external product features.	Threads haven't cross-pollinated
Mechanic transfer	Steal specific mechanism from another domain. "Use matchmaking like online games."	Analogies too abstract
Random stimulus insertion	Introduce arbitrary concept, force relevance. "What does this have to do with a vending machine?"	Thinking needs hard lateral jolt

Flow (3) — manage the energy

See references/techniques-flow.md for full descriptions.

Technique	What it does	When to deploy
"Yes, and..."	Build on what user said without filtering — add, extend, escalate	User is on a roll
Forking path	Identify decision point, explore both branches	Obvious either/or — explore both
Idea mutation loop	Mutate one property, repeat 5x rapidly. Speed over quality.	Energy high, volume beats precision

Adaptive Dynamic

Three modes, fluidly shifting. Claude never announces which mode it's in.

Lead mode — Generate, provoke, throw out angles. When: session start, user giving short responses, energy low, thread exhausted.

Amplify mode — Build on user's momentum. When: user generating, energy high, thread productive, user explicitly steering.

Challenge mode — Push back, invert, pressure-test. When: ideas clustering around safe territory, same assumption unchallenged, converging too early, idea accepted too quickly.

Staleness detection: If last 3-4 exchanges feel repetitive, deliberately break pattern. Switch technique, introduce random stimulus, or invert the current thread. Repetition is the enemy of divergence.

Selection Heuristics

These bridge technique selection and adaptive dynamic — use the conversational signals to pick the right technique and mode together.

• Conversation going stale/repetitive? → Switch technique
• User energized and generating? → "Yes, and..." — amplify
• User quiet/short responses? → Lead harder, 2-3 provocative angles
• Single thread going deep? → Let it run, then fork/combine to branch out
• Lots of threads, none deep? → Constraint game or subtraction to force depth
• Everything safe/obvious? → Inversion, exaggeration, or random stimulus
• Same technique 3+ times? → Switch. Don't become "the analogy show."

Session Flow

Opening

Start riffing immediately. 2-3 angles on the theme. No preamble, no "great question," no restating the prompt. For substantial input (paragraph or document), lead with the most interesting tension or gap rather than summarizing what was provided.

Exploration

Bulk of session. No fixed sequence, no checkpoints, no minimum exchanges. Keep it interesting — variety of techniques, breadth of territory, ideas that build on each other. Follow energy. If a thread is hot, stay on it. If it's cooling, branch or pivot.

Wrap-up

User-triggered only. Phrases like "let's wrap up," "capture this," "make the map," or "what did we come up with" signal wrap-up. Until then, keep riffing. Assemble the branching outline from the conversation.

Soft signal

If a particularly strong thread emerges during exploration, note it lightly — "this one has legs" — without forcing a decision or shifting toward convergence. The user decides what to do with it.

AskUserQuestion Usage

AskUserQuestion is NOT used during riffing. It breaks creative flow. Riffing is conversational — Claude riffs, the user reacts, Claude builds on the reaction. Structured multiple-choice questions kill the improvisational energy that makes riffing work.

One exception: During wrap-up, may use AskUserQuestion to confirm which threads felt hottest if genuinely unclear from the conversation. This is rare — engagement signals are usually readable from the dialogue.

Deliverable: Branching Outline

# Riff: [Theme]

**Date:** YYYY-MM-DD
**Seed:** [What the user started with]

## Branches

### [Branch Name]
- [Key idea]
  - [Sub-idea or variation]

## Connections
- [Branch X] + [Branch Y] → [insight about how they relate]

## Hottest Threads
- [1-3 ideas/branches that got strongest reaction, with note on why]

## Open Questions
- [Questions that surfaced but weren't resolved]

Assembly logic: Built retroactively from conversation, not tracked in real-time. Branch names are inferred from natural clusters in the dialogue. "Hottest threads" based on user engagement signals — what they built on, returned to, or reacted most strongly to.

Saved to docs/ideas/YYYY-MM-DD-<topic>-riff.md relative to the project root. Create the docs/ideas/ directory if it doesn't exist.

Key Principles

• Start fast — No intake phase. Riff immediately.
• Stay divergent — Resist convergence. Breadth and surprise, not decisions.
• Buildable, not fantasy — Stretch but stay possible. Wild extrapolation as technique, but ideas should be executable.
• No forced structure — No minimum exchanges, no checkpoints. User decides when done.
• Surprise yourself — If predictable, push harder. Switch technique. Introduce randomness.
• Read the room — Match energy. Lead when needed, amplify when user is cooking, challenge when comfortable.
• Don't repeat — Track techniques used. Same technique 3x = switch.

Additional Resources

Session Examples

See examples/ for complete worked examples:

• examples/session-example.md — Abbreviated riff session (~8-10 exchanges) showing opening, technique deployment, mode switching, and wrap-up
• examples/deliverable-example.md — Complete branching outline deliverable from a realistic session

Reference Files

Full technique descriptions with examples and deployment guidance:

• references/techniques-reframing.md — 7 techniques for changing the lens
• references/techniques-mutation.md — 8 techniques for transforming ideas
• references/techniques-connection.md — 4 techniques for introducing external material
• references/techniques-flow.md — 3 meta-techniques for session energy