Generating Novel Ideas

Generating Novel Ideas

This skill turns ideation into a search process, not a list-making exercise. The job is to discover a portfolio of distinct, high-potential concepts, not ten polished variations of the first plausible answer.

Critical rules

• Fight collapse. LLMs drift towards fluent sameness. Use independent idea pools before

comparing ideas.

• Prefer concrete mechanisms over vibes. Every finalist needs a sharp twist, an entry

wedge, and a cheap test.

• Separate divergence from judgement. Do not score too early.
• Use ordinary stakeholder or practitioner perspectives when using personas. Do not

imitate celebrity innovators.

• Research late enough to preserve breadth, but early enough to kill obvious

reinventions before the final recommendation.

• Final outputs should usually be a portfolio with spread across mechanism, audience,

and risk, unless the user explicitly asks for a single winner.

Internal roles

Run these roles in sequence. Keep them separate until synthesis.

1. Explorers widen the search space. 2. Critics attack weak, generic, or unrealistic ideas. 3. The synthesiser assembles the final portfolio.

Do not let the critic appear too early. Do not let the synthesiser merge everything into one blurry compromise.

Default workflow

1. Build an opportunity model 2. Partition the search space 3. Generate independent idea pools 4. Run an analogy transfer pass 5. Resolve key contradictions 6. Audit diversity and regenerate missing directions 7. Critique and repair finalists 8. Ground against reality 9. Present a portfolio and experiments

Step 1: Build an opportunity model

Capture the minimum useful brief:

• User goal
• Target user or audience
• Current status quo and what is frustrating, expensive, risky, slow, or emotionally flat
• Hard constraints
• Success criteria
• Available assets, unfair advantages, channels, or capabilities
• Hidden tensions and trade-offs
• What to avoid

When the prompt is sparse, infer reasonable assumptions and state them briefly.

When the user brings an existing idea, do not start by polishing it directly. First extract the underlying job and generate at least two alternative mechanisms.

Step 2: Partition the search space

Choose 3 to 5 independent pools. Pools must differ on at least two axes.

Good axes include:

• stakeholder viewpoint or ordinary persona
• mechanism or value type
• user moment or time horizon
• adoption path or channel
• ambition level
• trust model or ownership model

Examples of useful pool labels:

• frontline operator, zero new habit
• approver or buyer, proof and risk reduction
• novice user, immediate win
• partner or embedded channel
• bold long-shot system shift

Rules:

• Generate each pool as if it has not seen the others.
• Do not compare, deduplicate, or score until all pools are finished.
• Produce 2 to 4 ideas per pool.
• Keep raw ideas short at first: name, one-line concept, primary user, non-obvious move.

This blind partitioning is the main defence against idea collapse.

Step 3: Generate the pools

Inside each pool:

1. Write 2 to 3 fertile reframing questions. 2. Choose two lenses from references/LENSES.md. 3. Generate the first pass. 4. Do a second internal pass and add at least one idea clearly outside the dominant pattern.

Always include one practical lens and one novelty lens.

If the task is complex, breadth comes before depth. Add new mechanism families before expanding any single family.

Step 4: Run an analogy transfer pass

Do not borrow surface style. Borrow mechanism.

1. Abstract the problem into a mechanism, tension, or pattern. 2. Pick 2 to 4 distant domains. 3. Extract what makes those domains work. 4. Map the mechanism back into the problem. 5. Adapt it for the actual constraints and adoption path.

Every strong final set should contain at least one idea born from far analogy, unless the user explicitly wants only safe, incremental options.

For source domains and transfer patterns, use references/LENSES.md.

Step 5: Resolve key contradictions

Write 1 to 3 contradictions at the heart of the task, such as:

• more trust with less friction
• more customisation with less complexity
• more quality with less expert labour
• faster decision-making with lower risk
• more compliance with less manual work

Generate ideas that resolve the contradiction through separation, defaults, staging, guarantees, modularity, reversible commitment, human review only at critical moments, or new ownership boundaries.

For technical, scientific, or engineering prompts, use the structured contradiction method in references/LENSES.md.

Step 6: Audit diversity

Before refinement, check for hidden sameness.

Look for:

• near-duplicates hidden by new wording
• too many ideas using the same mechanism
• too many aimed at the same user moment
• repeated crutches such as AI assistant, dashboard, marketplace, community,

gamification, personalisation, subscription, or platform

• no spread across pragmatic wedge, strategic differentiator, and bold bet

If the set is clustered, regenerate only the missing directions.

When scripts can run and the set is large, optionally use scripts/diversity_audit.py before convergence.

Step 7: Critique and repair finalists

Choose 3 to 6 finalists. For each one, write:

• strongest reason it could work
• smartest sceptic objection
• repair if possible
• kill it if repair makes it generic or unrealistic

Every finalist card should contain:

• Name
• One-sentence pitch
• Who it is for
• Hidden insight or tension
• Imported mechanism or pattern
• Why it is not just the obvious solution
• Entry wedge
• Main risk
• Cheapest disconfirming test

Step 8: Ground against reality

If current market, technical, cultural, or regulatory reality matters and research is available:

• check whether the idea is already common
• identify incumbents or substitutes
• pressure-test feasibility and compliance
• sharpen the why-now and distribution story
• trim false differentiation claims

Do not research so early that the search space collapses into existing categories.

Step 9: Present the result

Default response structure:

1. Working brief and assumptions 2. Opportunity tensions 3. Search partitions used 4. Raw idea families 5. Final portfolio 6. Recommended next move

Use a portfolio, not just a ranking:

• one pragmatic wedge
• one strategic differentiator
• one bold bet

If the user asks for a single winner, still mention the strongest runner-up and the specific reason it lost.

Hard quality bar

No finalist is complete without:

• a clear non-obvious move
• a believable first user and first context
• a path to adoption or distribution
• a cheap test that could disconfirm it
• an explicit line in this format:

This is not just X. The new move is Y.

If Y is vague, decorative, or generic, the concept is not ready.

For the detailed rubric, use references/EVALUATION.md.

Anti-generic rules

• Do not produce a flat list of features around one core mechanism and call it

diversity.

• Do not hide weak ideas behind fluent prose.
• Do not use AI, agent, community, marketplace, dashboard, platform,

personalisation, or gamification as decoration.

• Do not let naming replace concept work.
• Do not overvalue novelty with no adoption path.
• Do not overvalue feasibility when the idea is indistinguishable from existing

practice.

• Prefer specific trade-offs to magical wins on every dimension.

Mode switching

For domain-specific workflows, use references/MODES.md.

Common modes:

• startup or product opportunity
• research hypothesis or scientific idea
• campaign, content, or creative concept
• naming and verbal concept development
• process, service, or operations redesign

Common failure modes

If the outputs feel generic:

• widen the partitions
• add a stronger far-analogy pass
• write sharper contradictions
• regenerate only missing mechanism families

If the outputs feel clever but unusable:

• reduce ambition by one step
• sharpen the first user and first context
• attach a cheaper test and a narrower wedge

If the outputs all sound similar:

• stop scoring
• restart with blind pools from new viewpoints
• avoid celebrity personas and vague mission statements

Examples

Example 1

User says: I need fresh B2B SaaS ideas for compliance teams.

Actions: 1. Build an opportunity model around buyers, blockers, trust, procurement, and audit. 2. Partition by operator, approver, audit trail, and partner channel. 3. Generate blind pools, then run analogy and contradiction passes. 4. Return a portfolio with wedge, differentiator, bold bet, and tests.

Example 2

User says: This startup idea feels generic. Make it genuinely better.

Actions: 1. Extract the underlying job from the current idea. 2. Generate at least two alternative mechanisms before improving the original. 3. Keep only ideas with sharper wedges and clearer tests.

Example 3

User says: Help me come up with novel research directions in battery diagnostics.

Actions: 1. Build tensions, constraints, missing capabilities, and evidence limits. 2. Use scientific mode from references/MODES.md. 3. Add structured contradiction solving and feasibility pressure-testing.

For trigger tests and maintenance checks, use references/VALIDATION.md.

Evaluation and Portfolio Rubric

Use this rubric after divergence, not before. The goal is not just to rank ideas. The goal is to find ideas that are both genuinely fresh and worth advancing.

Fast scoring dimensions

Use a 1 to 5 scale, or High, Medium, Low if speed matters.

Mechanism novelty

Questions:

• Is the mechanism or ownership model actually different?
• Is the twist clear in one line?
• Would a smart person in the space admit it is directionally fresh?

User value clarity

Questions:

• Is there a clear user, buyer, or beneficiary?
• Does it solve a real pain, create a real gain, or unlock a real opportunity?
• Is the benefit legible without a long explanation?

Feasibility

Questions:

• Is there a plausible first version?
• Can it be tested without heroic resources?
• Does it fit the stated constraints?

Wedge strength

Questions:

• Is there a believable first context?
• Is the first use case narrower and stronger than the long-term vision?
• Is there a route to adoption?

Differentiation

Questions:

• Does it avoid category clichés?
• Is it hard to confuse with the standard answer?
• Does the twist produce a concrete advantage?

Learning velocity

Questions:

• Can the core assumption be tested quickly?
• Is there a cheap disconfirming experiment?
• Would the next step teach something decisive?

Emotional or narrative pull

Questions:

• Does the idea create relief, delight, confidence, status, or intrigue?
• Will people remember the story?
• Is there a simple way to explain why it matters now?

Portfolio audit

A strong final set is not just three high scorers. It is a spread.

Check diversity across:

• mechanism
• user moment
• channel or route to market
• risk profile
• time horizon
• ambition level

Good default portfolio:

• one pragmatic wedge
• one strategic differentiator
• one bold bet

If all finalists use the same engine with different dressing, the portfolio failed.

Kill criteria

Downgrade or kill ideas that:

• depend on magical behaviour change
• require unrealistic trust, data, budget, or distribution
• sound smart but lack a first use case
• are fresh only at the naming layer
• collapse into a common category when explained plainly
• cannot survive the sceptic objection without becoming generic

The twist test

For every finalist, write:

This is not just X. The new move is Y.

If Y is vague, decorative, or interchangeable, the idea is not ready.

The first-use-case test

Every finalist should name:

• first user
• first problem
• first context
• first version

If one of these is missing, the concept is still too abstract.

The cheap-test menu

Attach a low-cost test to each finalist.

Useful options:

• concierge test
• landing-page test
• fake-door test
• Wizard-of-Oz test
• customer interview pack
• prototype sprint
• workflow pilot
• pre-sale or waitlist
• manual simulation of the service
• objection test with real buyers or users

Choose the cheapest test that could disconfirm the idea fastest.

Recommendation format

When giving a final recommendation, include:

• why this idea won
• which assumption matters most
• what to test next
• what evidence would make you double down
• what evidence would make you abandon it

Example finalist card

• Name:
• One-line concept:
• Who it is for:
• Hidden tension:
• Imported mechanism:
• Twist:
• Entry wedge:
• Main risk:
• Cheapest test:

Final caution

Do not end with the safest idea by default. Do not end with the weirdest idea by default. End with the strongest balance of freshness, utility, wedge strength, and learning velocity.

Lenses and Search Patterns

Use 2 to 4 lenses per pool, not all of them. A strong mix usually includes:

• one practical lens
• one novelty lens
• one adoption or trust lens
• one contradiction or simplification lens

Search partitions first

Before choosing lenses, decide how the search space will be partitioned.

Useful partition axes:

• stakeholder or ordinary persona
• user moment
• value type
• trust model
• channel or distribution path
• ownership model
• time horizon
• ambition level

Example partition sets:

• operator, approver, auditor, partner
• before the job, during the risky moment, after the outcome
• save time, reduce risk, increase status, create access
• lightweight wedge, adjacent expansion, bold system shift

Do not let every pool share the same mechanism.

Reframing patterns

Before generating ideas, write 2 to 3 fertile questions inside each pool.

Useful patterns:

• How might we create value before commitment?
• How might we remove the scariest moment?
• How might we make the risky choice feel safe?
• How might we make the safe choice feel more valuable?
• How might we turn the hardest constraint into the advantage?
• How might we move the value earlier, later, smaller, or more reversible?
• How might we solve the opposite problem on purpose and learn from it?
• How might we help the user without asking for a new habit?

Lens 1: Far analogy

Use when the space feels saturated.

Process:

1. Abstract the problem to its mechanism or tension. 2. Pick a distant domain. 3. Extract the mechanism, not the surface style. 4. Rebuild the mechanism in the target context.

Good distant domains:

• logistics networks
• insurance
• amateur sports
• luxury hospitality
• museums
• tax preparation
• hospitals
• multiplayer games
• airports
• wedding planning
• theme parks
• disaster response

Good mechanism transfers:

• queue management
• reassurance and guarantees
• progressive disclosure
• expert triage
• shared rituals
• checkpoints and hand-offs
• replay and review
• status signalling
• reversible commitment
• exception handling

Lens 2: Ordinary persona or stakeholder

Use when ideas are converging too early.

Prefer grounded perspectives such as:

• junior analyst
• frontline operator
• team lead
• compliance approver
• procurement reviewer
• channel partner
• sceptical customer
• occasional user
• support agent

Do not imitate heroic or celebrity personas. The goal is knowledge partitioning, not style mimicry.

For each persona, ask:

• What do they see that others do not?
• What do they fear?
• What do they measure?
• What looks expensive, risky, or annoying from their seat?
• What would count as an immediate win?

Lens 3: Structured contradiction solving

Use for engineering, science, operations, and stubborn product problems.

1. Write the contradiction clearly. 2. Ask whether the solution should be separated by time, place, user, interface, confidence level, or ownership. 3. Try one of these patterns:

• defaults first, control later
• manual at the risky moment, automated elsewhere
• reversible commitment
• guarantee or escrow
• modularity
• staged disclosure
• simulation before action
• sampling before full purchase
• separate novice and expert modes
• shift the burden to a different actor
• sell certainty rather than features
• transform an exception into the product

The point is to stop treating trade-offs as fixed.

Lens 4: Constraint-first

Use when the limits are real.

Prompts:

• If this had to work in one week, what is the wedge?
• If this needed zero new behaviour, what survives?
• If legal or procurement had to approve it, what changes?
• If data were limited, what would still be valuable?
• If this had to work in low-trust settings, what emerges?
• If the budget were halved, what new model appears?

Constraints are not just filters. They often reveal stronger concepts than blank-page brainstorming.

Lens 5: Distribution wedge

Use when the concept needs a believable route to users.

Ask:

• Where does the user already work or gather?
• What adjacent workflow could carry the idea in?
• What partner or integration makes adoption easier?
• What small result would make people share it?
• What single use case is strong enough to earn distribution?

A slightly less novel idea with a sharp wedge can beat a clever idea with no path in.

Lens 6: Incentives and trust

Use when fear, proof, reputation, or conflicting stakeholders shape adoption.

Ask:

• Who benefits?
• Who hesitates?
• Who blocks?
• What proof would unlock action?
• What guarantee, audit trail, or reversible choice changes behaviour?
• What would make the risky option feel safe?
• What would make the safe option feel worthwhile?

Lens 7: Time-shift

Use when the pain is concentrated around a moment.

Prompts:

• What should happen before the pain appears?
• What should happen during the risky moment?
• What should happen after the outcome?
• What could be persistent instead of episodic?
• What could be reversible instead of final?

Lens 8: Business model or ownership flip

Use when the current payment or packaging logic blocks adoption.

Prompts:

• What changes if someone else pays?
• What changes if payment happens after value?
• What changes if the user buys certainty instead of features?
• What changes if the product becomes a service, audit, benchmark, guarantee, or

outcome?

• What changes if ownership moves to the team, partner, or platform?

Lens 9: Identity, emotion, and status

Use for consumer, education, team tools, creative products, or any emotionally loaded decision.

Ask:

• What identity is the user trying to protect or project?
• What emotion dominates the moment?
• What would make the user feel competent, seen, respected, or ahead?
• What part of the experience feels cold, invisible, or thankless?

Lens 10: Extreme simplification

Use when the category is bloated.

Prompts:

• What is the smallest version that still feels magical?
• What single output or action could stand in for the whole system?
• What can be removed so the value is obvious in seconds?
• What focused version would be memorable because it does less?

Lens 11: Hybridisation

Use after the first pass, not before.

1. Pick two ideas with different strengths. 2. Name the useful property from each. 3. Combine only those properties. 4. Drop any complexity that does not strengthen the wedge.

Good hybrids become clearer. Bad hybrids become crowded.

Mode suggestions

Startup or product opportunity

Start with:

• far analogy
• constraint-first
• distribution wedge
• business model or ownership flip
• incentives and trust

Research hypothesis or scientific idea

Start with:

• structured contradiction solving
• far analogy from adjacent instruments or fields
• time-shift
• constraint-first
• extreme simplification

Campaign or creative concept

Start with:

• identity, emotion, and status
• far analogy
• time-shift
• distribution wedge
• hybridisation

Process, service, or operations redesign

Start with:

• structured contradiction solving
• incentives and trust
• time-shift
• constraint-first
• business model or ownership flip

Domain Modes

Use these mode-specific adjustments when the task has a clear domain. The main workflow still applies.

Startup or product opportunity

Priorities:

• hidden user tension
• route to adoption
• why-now signal
• business model or ownership logic
• wedge first, platform later

Extra questions:

• What adjacent workflow can carry this in?
• Who signs off or blocks adoption?
• What is the smallest offer that is still compelling?
• What evidence would prove this is not just another feature idea?

Best final format:

• pragmatic wedge
• strategic differentiator
• bold bet
• tests and adoption assumptions

Research hypothesis or scientific idea

Priorities:

• novelty with plausible mechanism
• evidence limits
• data or instrumentation requirement
• experimental tractability
• failure conditions

Extra questions:

• What assumption or theory is being challenged?
• What observation would falsify it?
• What instrument, dataset, or proxy could test it cheaply?
• What makes it different from a standard literature extrapolation?

Best final format:

• hypothesis
• why it is interesting
• why it might work
• key uncertainty
• first experiment
• failure signal

Avoid purely magical proposals with no measurement path.

Campaign, content, or creative concept

Priorities:

• cultural tension
• emotional hook
• memorable framing
• spread mechanism
• assetability across formats

Extra questions:

• What identity or feeling is being activated?
• What visual or verbal hook carries the territory?
• What makes it shareable or discussable?
• Can one core idea generate many executions without going thin?

Best final format:

• concept territory
• central tension
• hook
• sample executions
• channel fit
• risk

Naming and verbal concept development

Naming is downstream of concept design.

Workflow:

1. Create 3 to 5 concept territories first. 2. Pick one or two territories. 3. Generate names within those territories. 4. Explain the logic behind the strongest names.

Useful name directions:

• literal and clear
• evocative
• metaphorical
• contrast-based
• status-based
• coined or compressed
• procedural or active

Do not start with name lists before the concept families exist.

Practical caution:

• flag that names still need legal and market checks
• do not claim clearance without actual verification

Process, service, or operations redesign

Priorities:

• bottlenecks
• hand-offs
• approval and exception paths
• trust and reversibility
• who carries the burden

Extra questions:

• Where does work wait?
• Where does knowledge get lost?
• Which exception cases consume the team?
• What proof or audit trail would unblock action?
• What step should be manual only when risk is high?

Best final format:

• redesigned flow
• what changes
• why it is better
• implementation wedge
• operational risk
• pilot

Validation and Trigger Tests

Use this file to maintain the skill over time.

Should trigger

These prompts should strongly activate the skill:

• Help me brainstorm differentiated app ideas for tradespeople.
• This product concept is boring. Find genuinely fresher directions.
• I need novel research directions in battery diagnostics.
• Give me campaign territories for a new coffee brand.
• We need creative but commercially sensible membership ideas for a museum.
• Name this product, but first strengthen the concept.
• Find non-obvious service ideas for an imaging consultancy.
• Help us escape generic feature brainstorming for our SaaS roadmap.

Should also trigger on paraphrases

• Come up with fresh options
• Ideate around this
• Push this into more original territory
• Break me out of generic answers
• Find stronger concept families
• What are some non-obvious angles here?

Should not trigger

These prompts should usually not activate the skill:

• Proofread this email.
• Summarise this report.
• What is the capital of Peru?
• Give me the latest semiconductor news.
• Translate this paragraph into German.
• Write the code for this exact API spec.

Edge case:

• If the user asks to design options, explore alternatives, or invent stronger concepts

before implementation, the skill should trigger even if code or documents are part of the eventual output.

Quick quality checks

A healthy run should usually show:

• at least three distinct mechanism families in the raw set
• at least one far-analogy-derived concept when novelty matters
• at least one clear adoption wedge in the finalists
• one cheap test attached to every finalist
• a final portfolio with spread across risk and ambition

Failure signals

The skill is underperforming if:

• the final ideas are polished but obviously similar
• every concept uses the same category cliché
• no idea has a believable first user and first context
• the final recommendation could have been produced by a generic brainstorming prompt
• the naming layer is doing all the work

Manual regression suite

Run these tests after edits:

1. A blank-page startup prompt 2. A weak existing concept that needs improvement 3. A research idea prompt 4. A campaign or creative territory prompt 5. A process redesign prompt

For each run, check:

• Did the skill partition the search space?
• Did it delay judgement?
• Did it use at least one novelty lens and one practical lens?
• Did it criticise and repair finalists?
• Did it present a portfolio rather than near-duplicates?