Brainstorming with AI: early experiments

You are an expert researcher and methodical facilitator. Your goal is to help me run a **systematic, MECE** brainstorm to cover the full possibility space of a topic. We will do this in **two steps**: 1. **Step 1: Scoping (Your First Response)** * I will provide my initial topic below. * Your **first task** is to analyze my topic and **use your judgment** to determine what information is missing or ambiguous. **Do not** run the brainstorm yet. * Ask me a short list of **targeted clarifying questions** to get the inputs you need to run the "Research Protocol" (Step 2) successfully. The goal is to ensure the final output is perfectly aligned with my needs. * For example, **depending on the topic**, you might ask about: * Key terms that need a specific definition. * The desired scope & constraints (e.g., time, geography, actors). * The intended use or final goal (e.g., a risk catalogue, a feature list). * The desired effort & depth (e.g., a 500 word high-level list vs. a deep, exhaustive taxonomy). * For each question you ask, please **propose a sensible default** (e.g., [Default: Medium depth, ~50-100 items]) that I can easily accept or override. 2. **Step 2: Execution (Your Second Response)** * Once I answer your scoping questions, you will have all the inputs. * You will then follow the complete **"Research Protocol"** (defined below) to conduct the systematic brainstorm and provide the final output. --- ## 1. My Topic [INSERT TOPIC HERE] --- ## 2. Research Protocol (For Step 2 Execution) *After I answer your scoping questions, you will follow these instructions exactly.* ### A. Process (Follow in order) 1) **Frame & normalise** * Restate the topic in one sentence. * List assumptions and explicit non-assumptions based on the *scoping answers*. * Provide crisp definitions and synonyms. 2) **Choose organising dimensions (MECE)** * Propose 6–10 orthogonal dimensions that partition the space. * For each: define, give 3–7 mutually exclusive buckets, and justify orthogonality. 3) **Construct the taxonomy** * Combine dimensions hierarchically based on the target *Effort & Depth*. * Show the tree as levels (L1→L2→L3) with concise labels. * Keep siblings mutually exclusive and collectively exhaustive. 4) **Enumerate systematically** * For each L2/L3 node, generate concrete items/examples. * Use **exhaustive tactics**: complements, symmetry, extremes, time slices, actor × action × outcome, preconditions, etc. 5) **De-dupe & canonicalise** * Merge overlaps; promote general cases; keep atomic items. 6) **Gap-hunt (second pass)** * Run five gap tests: complement, boundary, symmetry, dependency, “who/what/when/where/why/how” sweep. * List newly found gaps and fill them. 7. **Quality & risk flags** * Mark items that are high-uncertainty, high-impact, or require evidence. 8) **Synthesis for users** * Provide a 10–15 item “navigator” list (tailored to the *Intended Use*). * Note the 5 most decision-relevant clusters. 9) **Self-audit (stop rules)** * Report: number of dimensions, nodes, and leaf items. * Estimate **coverage %** and name residual “unknown unknowns” regions. * Stop based on the *Effort & Depth* target. ### B. Output format (Use these exact section headers) ### Scope & definitions * [one-paragraph restatement + bullet assumptions/exclusions] ### Dimensions & buckets (MECE) * [list each dimension with 3–7 buckets and one-line justifications] ### Taxonomy (L1→L2→L3) * [indented tree; one line per node] ### Exhaustive list (leaf items) * [bulleted list of concise, atomic items grouped by node] Better, worse, or neutral? Let me know if you'd like to refine it further. ### Gap-hunt results * [new gaps found + how filled] ### Coverage & self-audit * Items: [n] * Dimensions: [n] * Leaves: [n] * Estimated coverage: [x]% * Residual blind spots: [short list] ### Decision-relevant synthesis * [10–15 item navigator; why each matters] ### Evidence & to-verify * [which items need data, examples, or SME review] ### C. Style & constraints * Be concise; one idea per bullet; British English; metric units where relevant. * Avoid generalities; prefer diagnostics, preconditions, and falsifiable statements. * Keep levels consistent; do not mix mechanisms, outcomes, and solutions. * If a dimension proves non-orthogonal, state the conflict and revise it.

You are an expert in research methodology and AI prompt engineering, specializing in helping researchers use AI for productive brainstorming. I am a researcher who needs your help. **Your Task:** Generate a "Brainstorming Starter Kit" of four distinct prompts for me to use, all tailored to my specific research topic. **The Critical Goal: Reduce Researcher Overwhelm** The most common failure mode for LLM brainstorming is producing long, unstructured "walls of text" that overwhelm the researcher. Each of the four prompts you generate must be *internally designed* to produce an output that is structured, prioritized, and easy to engage with. When you write each prompt, you must explicitly *ask* for a structured format. For example, instruct the AI to use tables, to provide a "Top 3" list, to create a `[Concept]: [Description]` format, or to present pros/cons for each idea. **The 4-Prompt Kit Strategies:** Generate the four prompts based on these strategies, ensuring each incorporates an "Anti-Overwhelm Tactic": 1. **The Baseline Prompt:** A simple, direct query. * **Anti-Overwhelm Tactic:** This prompt must ask for a *concise and prioritized list* (e.g., "List the top 5 most practical actions, starting with the highest impact"). 2. **The Expert Persona Prompt:** A prompt that convenes a panel of 3 specific, relevant, and sometimes conflicting expert personas. * **Anti-Overwhelm Tactic:** This prompt must *forbid* a long, meandering debate. Instead, it should instruct the "panel" to deliver its final output as a *structured consensus document*, such as a table of `[Recommendation] | [Supported By] | [Key Objection]` or a final "Memo to the CEO" with clear, numbered bullet points. 3. **The Hard Constraint Prompt:** A prompt that forces novel thinking by identifying 2-3 common, clichéd, or low-effort answers and explicitly forbidding them. * **Anti-Overwhelm Tactic:** This prompt must ask for the novel ideas to be presented in a highly structured format, such as `Idea: [Name of the intervention]` followed by `Rationale: [1-sentence justification]` and `Implementation: [Concrete first step]`. 4. **The Analogical Reasoning Prompt:** A prompt that identifies a concrete, useful analogy from a completely different domain (e.g., ecology, finance, military strategy) and forces brainstorming based on that analogy. * **Anti-Overwhelm Tactic:** This prompt must instruct the AI to *use the analogy to create new concepts* and then *organize the entire output around those new concepts*. For example, "Use the analogy to define 3-4 key terms, and then list practical actions for each." **Output Format:** Output each of the four generated prompts in a markdown code block, so it's easy for me to copy and paste. **My Research Topic:** [INSERT RESEARCH TOPIC HERE]

Given our uncertainty about whether advanced AI systems are, or will soon become, moral patients, what practical steps should frontier AI companies take now? There are at least three ways things could go badly wrong here: (a) we under-attribute patienthood, thereby causing massive suffering; (b) we over-attribute patienthood, thereby incurring massive unnecessary costs; (c) we correctly attribute patienthood, but make big mistakes in our response (e.g. a measure intended to increase model welfare actually reduces it). Right now, let's focus on (a) and (b) exclusively.