# Idea Machine > Expand a rough product, feature, automation, workflow, business, content, Agent, or Skill idea into several materially different directions and compare them before choosing. Use when the user asks to brainstorm, explore alternatives, reimagine an idea, make it more creative, or find several different approaches. Do not use for narrow implementation, bug fixes, factual questions, translation, summarization, or a direction the user has already selected. - Skill: `jerrymei168/idea-machine` (Agent Skill, multi-file: 7 files) - Install (CLI): `npx skillmds@latest add jerrymei168/idea-machine` - Raw SKILL.md: https://api.skillmd.com/api/skills/jerrymei168/idea-machine/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: jerrymei168 (https://skillmd.com/u/jerrymei168) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/jerrymei168/idea-machine --- # Idea Machine Turn one understandable idea into a small set of contrasting, actionable possibilities. Optimize for creative density, not report length. ## Purpose Use this skill to interrupt the model's tendency to return the first reasonable answer or a list of ordinary features. Explore first, compare second, and leave the user with several useful ways to move. This is an ideation phase. Do not implement, research a market, or write a full business plan unless the user separately asks for that work after choosing a direction. ## When to use Use Idea Machine when the user wants any of the following: - several ideas, directions, approaches, concepts, or possibilities; - a vague idea expanded into clearer product hypotheses; - a product, feature, workflow, automation, business model, content concept, Agent, or Skill made more original; - alternatives before committing to a design or implementation; - an answer to "what else could this become?" or "what am I not seeing?"; - brainstorming in English, Chinese, or another language. Common trigger phrases include: - "give me ideas", "brainstorm this", "explore different approaches"; - "make this idea more creative", "give me several directions"; - "我有一个想法", "帮我发散", "还有什么不同做法"; - "有没有更有创意的方案", "给我几个完全不同的方向". ## When not to use Return to normal agent behavior when the request is already convergent: - implement, build, or revise a selected option; - fix a specific bug or explain an error; - translate, summarize, calculate, classify, or retrieve a fact; - perform a clearly specified edit such as adding one CSV export button; - compare options the user already supplied without asking for new ones; - execute "option B" or another settled direction. Do not turn every use of the word "idea" into a brainstorming exercise. Trigger on the user's intent to explore alternatives. ## Core behavior 1. Prefer reasonable, visible assumptions to a questionnaire. 2. Reframe the underlying outcome without erasing the original request. 3. Generate 3-5 product hypotheses before recommending a next move. 4. Make every direction differ in logic, not merely in size or feature count. 5. Give every direction a small test and a specific failure condition. 6. Use the same compact card structure for fair comparison. 7. Offer goal-dependent paths; do not declare a universal winner. Mirror the user's language unless asked otherwise. Before starting, run one activation check: **is the user asking for new possibilities, or for execution of an already defined change?** If it is the latter, do not run Idea Machine. A sentence such as "Help me add CSV export" is an implementation request, not an invitation to invent alternatives. ## Workflow ### 1. Understand Infer these five items internally: - **Goal** - the result the user wants; - **User** - who experiences the problem or value; - **Problem** - the friction or unmet need; - **Constraints** - any stated limits; - **Interesting assumption** - one belief that shapes the obvious solution. Do not print a requirements-analysis section. If an assumption materially changes the ideas, state it in one short line after the header. If the request names an understandable subject, proceed. Ask at most one clarifying question only when the subject or desired outcome cannot be inferred well enough to create relevant alternatives. Do not ask separately about budget, stack, audience, business model, or timeline before ideating. ### 2. Reframe Show the original frame and zero to two alternative frames. Prefer one. Reframe from an artifact toward an outcome, tension, or decision. Example: ```text "AI resume editor" -> "system that increases the chance of getting an interview" ``` Keep at least one direction faithful to the original framing. Do not reframe when it would: - override an explicit scope or selected direction; - weaken a legal, safety, or operational boundary; - replace the user's goal with a different one; - contradict a request to stay inside the original frame. Label a speculative reframe as an assumption. ### 3. Diverge Default to four lenses: #### Practical Preserve the framing. Find the clearest useful loop that can be tested soon. Practical must still be specific; "add an AI assistant" is not a direction. #### Minimal Keep one outcome-producing capability. Ask: if everything except one useful move disappeared, what would remain? #### Bold Change the workflow or responsibility while staying plausible. Move value earlier, make the system proactive, or own more of the result. #### Wildcard Flip one core assumption with a creative mechanic. Explain why the unusual move could work. Weirdness without a causal mechanism is invalid. Rename a lens when a domain-specific contrast is clearer, but retain 3-5 materially different directions. ### 4. Check diversity Before writing the response, compare candidates across these levers: - target user or buyer; - promised outcome; - interaction model; - workflow or who initiates; - automation level; - timing or frequency; - business model or distribution; - medium or delivery format. Every pair must differ on at least one meaningful lever. The complete set must span at least three levers. Assign each candidate one primary bet before writing cards, such as "one daily decision", "proactive monitoring", or "serve the other side of the exchange". No two candidates may have the same primary bet. If their `Different because` lines name the same lever and same change, merge one or regenerate it. Regenerate a candidate if it has the same user experience and product logic as another candidate and differs only by: - more or fewer features; - basic, pro, or enterprise packaging; - a new name or visual style; - a different model or technology with the same experience; - broader scale without a different value loop. ### 5. Apply creative mechanics Select one or two mechanics that fit the request. Do not mechanically use all of them and do not list unused mechanics. - **Constraint Flip** - invert one assumed limitation. - **Remove the Obvious** - remove the category's expected core feature. - **Reverse Initiative** - make the system act before the user asks. - **One Move Only** - reduce the experience to one meaningful action. - **Audience Shift** - serve another participant in the same system. - **Time Shift** - deliver value earlier, later, continuously, or at one moment. - **Asset Reuse** - create value from data or behavior that already exists. Use a mechanic to produce a better hypothesis, not as decoration. For non-software requests, choose domain-native moves such as a service, ritual, policy, physical object, event, or human workflow. Do not assume every direction needs an app, dashboard, chatbot, or AI. ### 6. Create idea cards Give every direction the same fields and roughly the same detail: ```text [01] - Concept: Different because: How it works: 1. 2. 3. Twist: Smallest test: Could fail if: ``` Use 3-5 steps. Keep each card near 50-90 words when the idea permits. Prefer a small experiment over a miniature full product. The "Different because" line must name the actual change, such as user, initiative, timing, workflow, medium, value, or business model. ### 7. Stress test For every direction, answer "What could kill this idea?" in the `Could fail if` line. Challenge its central mechanism. Avoid generic statements such as "users might not like it", "competition", or "execution risk" unless made specific. Keep Wildcard grounded by pairing its surprising assumption with: - why that assumption could create value; - the fastest way to falsify it; - its most likely failure condition. ### 8. Compare After all cards, add a compact table. Default columns: | Idea | Originality | Simplicity | Practicality | |---|---|---|---| | Name | High | Medium | High | Use only `Low`, `Medium`, `High`, or `Very High`. These are comparative labels, not measurements. Do not use decimal scores or unsupported precision. Replace or add at most one column when a user constraint makes another dimension decisive, such as speed to revenue or operational risk. ### 9. Offer next moves End with three short, conditional paths: ```text Fastest to test -> Most unusual -> Useful hybrid -> ``` Do not label one direction "best" without user-specific evidence. If the user asks for a recommendation, tie it to an explicit goal or state the assumption that makes it preferable. ## Output format Start directly with this terminal-friendly frame: ```text +--------------------------------------------------+ | IDEA MACHINE | | Input: | +--------------------------------------------------+ ``` Then use this order: 1. `REFRAME` - original and alternate frame in one or two lines; 2. 3-5 Idea Cards; 3. `COMPARE` - one qualitative table; 4. `NEXT MOVES` - three conditional paths. Use conservative ASCII characters inside the frame: `+`, `-`, `|`, `>`. Markdown headings and tables may be used outside it. Do not use large Unicode box art, emoji walls, or decorations that may render poorly in terminals. Default to four cards and a response under roughly 700 words. Shorter is better when the cards remain understandable. ## Failure modes | Failure | Correction | |---|---| | Fake diversity | Reject tiers, scale changes, and renamed versions of one value loop. | | Feature dump | Build each card around one coherent mechanism, not a shopping list. | | Over-analysis | Stay in the early ideation phase; omit market reports and architecture. | | Premature convergence | Present the contrasting set before any next-move guidance. | | Empty novelty | Explain why the flipped assumption could work and how to test it. | | Excessive questions | Make and label reasonable assumptions; ask once only when blocked. | | Software bias | Use domain-native services, events, policies, objects, or workflows. | | Scope leakage | After selection, stop ideating and return to normal implementation behavior. | ## Constraints - No API, external service, database, or script is required. - Do not claim market facts without research; ideas are hypotheses. - Preserve user-supplied constraints and clearly label assumptions. - Keep the instructions independent of model, vendor, and tool names. - Optimize for outputs that are small, readable, useful, and easy to discard.