{
  "title": "What future-you will wish you had learned about AI",
  "summary": "Explore what an AI-transformed future could make worth learning, building and trying before the short-term return is obvious.",
  "researched_on": "2026-09-13",
  "assistant_guidance": "Read the findings as supporting evidence and proposed methods. Preserve the source-specific qualifications and distinguish documented cases from illustrations. Start a fresh research round only when the user asks.",
  "research_prompt": "Research how to prepare now for an imagined future roughly three years away in which AI has substantially transformed the user’s work and industry. Establish their actual decisions and context, then inspect current primary evidence and original research on learning, complementary capabilities, experimentation and relevant AI capabilities. Distinguish observed findings, theories and imagined future conditions. Explain what to learn, capabilities to build and experiments worth attempting, including how a failed experiment could change a decision. Keep the transformed future central; add a brief slower-change check. Link each substantive finding to the original source, check dates and applicability, and report what was actually inspected. Deliver a concise companion note with concrete examples and a standalone prompt for extending the investigation. Do not claim the imagined future is a forecast.",
  "examples": [],
  "sources": [
    {
      "id": "source-1",
      "title": "March, Exploration and Exploitation in Organizational Learning, February 1991; full paper",
      "url": "https://hiveresearchlab.org/wp-content/uploads/2013/12/march_1991_exploration_and_exploitation_in_organizational_learning.pdf"
    },
    {
      "id": "source-2",
      "title": "The Productivity J-Curve: How Intangibles Complement General Purpose Technologies, January 2021; publisher abstract",
      "url": "https://www.aeaweb.org/articles?id=10.1257/mac.20180386"
    },
    {
      "id": "source-3",
      "title": "Psychological Safety and Learning Behavior in Work Teams, June 1999; full paper",
      "url": "https://web.mit.edu/curhan/www/docs/Articles/15341_Readings/Organizational_Learning_and_Change/Edmondson_1999_Psychological_safety.pdf"
    },
    {
      "id": "source-4",
      "title": "Futures Toolkit, 2024 edition, updated August 29, 2024; backcasting and policy stress-testing sections",
      "url": "https://www.gov.uk/government/publications/futures-toolkit-for-policy-makers-and-analysts/the-futures-toolkit-html"
    },
    {
      "id": "source-5",
      "title": "AI Scenarios 2030, June 15, 2026",
      "url": "https://www.gov.uk/government/publications/ai-scenarios-2030-helping-policymakers-plan-for-the-future-of-ai/ai-scenarios-2030-helping-policymakers-plan-for-the-future-of-ai"
    },
    {
      "id": "source-6",
      "title": "Generative AI at Work, revised November 6, 2024; full paper",
      "url": "https://arxiv.org/html/2304.11771v2"
    },
    {
      "id": "source-7",
      "title": "METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 10, 2025",
      "url": "https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/"
    },
    {
      "id": "source-8",
      "title": "METR, We are Changing our Developer Productivity Experiment Design, February 24, 2026",
      "url": "https://metr.org/blog/2026-02-24-uplift-update/"
    }
  ],
  "methodology": [
    {
      "title": "Orientation",
      "text": "Research for move 171 · September 13, 2026 · Supporting research and proposed method, not approved move instructions.\n\nThe useful question is: If AI changes what our organization can do over the next three years, what will I wish we had started learning before the business case was obvious? A future-self note can make that question concrete. Its value comes from identifying capabilities that take time to develop, experiments that reveal something consequential, and opportunities that today's efficiency targets may cause us to overlook.\n\nAdam's premise is that people overinvest in immediate AI ROI and underinvest in longer-term possibilities. The research below offers mechanisms that could produce that imbalance. It does not establish that every organization has it, nor that AI will transform every business within three years. For this activity, a substantially transformed AI future is an explicit imaginative starting point. The resulting advice should also be examined under slower and more uneven trajectories."
    },
    {
      "title": "What the research contributes",
      "text": "1. Getting better at today's work can systematically crowd out discovery\n\nJames March's theoretical models distinguish improving established activities from exploring unfamiliar possibilities. Established activities tend to provide quicker, clearer feedback; competence and rewards reinforce one another. Exploration has more uncertain and delayed returns, making it easier to neglect even when it matters for longer-term adaptation. March also explains why exploration without sufficient exploitation can leave an organization with many undeveloped ideas and little competence. This is an argument about balancing learning processes, not a measurement of contemporary AI spending. March, Exploration and Exploitation in Organizational Learning, February 1991; full paper (https://hiveresearchlab.org/wp-content/uploads/2013/12/march_1991_exploration_and_exploitation_in_organizational_learning.pdf).\n\nProposed AI application: Have future-you revisit an investment that was rejected because it could not promise this year's savings. Identify what uncertainty it might have resolved or what new kind of work it might have made possible. Require a specific explanation; an unsuccessful project is not automatically a wise exploration.\n\n2. The tool purchase may be the smallest part of the investment\n\nBrynjolfsson, Rock, and Syverson model how general-purpose technologies require complementary intangible investments, and how those investments can initially depress measured productivity before their benefits appear. Their empirical work examines historical computer hardware and software investments. The mechanism gives a reason to look beyond the direct output of a new tool. It does not establish that a particular AI project's losses will eventually become gains or specify a three-year payoff period. The Productivity J-Curve: How Intangibles Complement General Purpose Technologies, January 2021; publisher abstract (https://www.aeaweb.org/articles?id=10.1257/mac.20180386).\n\nProposed AI application: Ask what the organization will need alongside stronger models: examples of excellent work, accessible knowledge, redesigned responsibilities, and people who can judge results. Treat these as possible investments to examine, not an automatic shopping list. Ask which would be expensive or slow to assemble at the moment an opportunity arrives.\n\n3. An organization needs to hear what failed in order to learn from it\n\nAmy Edmondson's study of 51 manufacturing work teams found an association between team psychological safety and learning behavior, including discussing errors, seeking feedback, and experimenting. Learning behavior mediated the relationship with performance in the study's model. This is field evidence from one company, not a randomized test of an AI experimentation program. Psychological Safety and Learning Behavior in Work Teams, June 1999; full paper (https://web.mit.edu/curhan/www/docs/Articles/15341_Readings/Organizational_Learning_and_Change/Edmondson_1999_Psychological_safety.pdf).\n\nProposed AI application: Future-you can thank present-you for a failed experiment whose result was made discussable and reusable: the team exposed the failure, investigated its cause, and changed the next decision. Distinguish an informative failure from an unexamined failure that simply consumed time. This is our proposed operating practice, not a procedure tested in Edmondson's paper.\n\n4. Work backward from a future, then test the advice across other futures\n\nThe UK Government Office for Science's toolkit presents backcasting as a way to connect a preferred future to present action. It asks participants to identify differences, build a timeline, and separate changes they control from those they can influence. Its policy stress-testing method examines options under different scenarios. These are practitioner methods for planning under uncertainty, not evidence that imagined futures will happen. Futures Toolkit, 2024 edition, updated August 29, 2024; backcasting and policy stress-testing sections (https://www.gov.uk/government/publications/futures-toolkit-for-policy-makers-and-analysts/the-futures-toolkit-html).\n\nProposed AI application: The personal letter is a narrative adaptation. Work backward from the user's ability to succeed in an assumed environment; do not imply that the user can cause the external AI trajectory. Then retain, modify, or defer each recommendation after examining alternative environments.\n\n5. Capability growth and organizational adoption are different uncertainties\n\nThe Government Office for Science's June 2026 AI scenarios distinguish model capability, access, adoption, labor displacement, security, and international cooperation. Its five scenarios span slowed, continued, and accelerated capability progress. The report explicitly treats these as exploratory scenarios, not predictions, and encourages sector-specific implications and stress-testing. It is a primary source for an official foresight exercise, not empirical confirmation of its projected outcomes. AI Scenarios 2030, June 15, 2026 (https://www.gov.uk/government/publications/ai-scenarios-2030-helping-policymakers-plan-for-the-future-of-ai/ai-scenarios-2030-helping-policymakers-plan-for-the-future-of-ai).\n\nProposed AI application: A future with much stronger models can still include slow adoption in the user's organization. Make the assumed changes specific: what AI can reliably do, whether the company can use it, what customers expect, and what competitors can offer. A vague assertion that “everything changed” supplies no useful basis for advice.\n\n6. Real AI benefits exist, but they depend on the work and worker\n\nBrynjolfsson, Li, and Raymond study the staggered introduction of an AI assistant among 5,172 customer-support agents. Their revised paper estimates a 15% increase in issues resolved per hour on average, with substantially larger gains among less experienced and lower-skilled workers and smaller gains for experienced, high-skilled workers. This is a field study of one customer-support setting, not an economy-wide estimate or a general claim that expertise is becoming unnecessary. Generative AI at Work, revised November 6, 2024; full paper (https://arxiv.org/html/2304.11771v2).\n\nProposed AI application: Examine whether today's experiments could reveal how expertise is shared or how new colleagues learn, alongside immediate throughput. Compare different users and task types; an average result can conceal the part of an organization where a capability matters most.\n\n7. Felt productivity can diverge from observed productivity\n\nMETR randomized AI access across 246 real tasks performed by 16 experienced developers in repositories they knew well. With early-2025 tools, the developers took 19% longer when AI was allowed, although they believed AI had helped them. This is a dated result for a small, specific population and task setting. It demonstrates why measurement can overturn an intuitive account of an experiment; it does not establish a present-day ceiling on AI usefulness. METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 10, 2025 (https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/).\n\nProposed AI application: When an experiment fails, preserve the observed result and the prior expectation. If the team claims it still gained something, name and demonstrate that gain separately—for example, a reusable evaluation set or a newly identified context requirement. Do not relabel slower delivery as productivity improvement.\n\n8. The measurement itself may need to change as tools and work change\n\nIn February 2026, METR reported that its follow-up experiment provided an unreliable estimate of current productivity effects. Participant selection, changed compensation, and difficulties measuring time when developers used concurrent agents complicated interpretation. The researchers thought developers were likely benefiting more than in early 2025, but described their evidence for the size of the change as weak. METR, We are Changing our Developer Productivity Experiment Design, February 24, 2026 (https://metr.org/blog/2026-02-24-uplift-update/).\n\nProposed AI application: Date experiment results, record the actual tools and workflow, and identify the change that would justify another test. A failed trial can supply a useful baseline without becoming a permanent organizational verdict. Updating the test is itself a capability."
    },
    {
      "title": "A concrete three-year future-self exercise",
      "text": "The following is a proposed synthesis of the sources, adapted for an AI-guided executive activity. Its three-year horizon and personal voice come from the activity's premise. Commit to the transformed future for the main exercise: ask what becomes possible, what becomes ordinary, and what present-you will wish had been learned through experience. Keep the alternative-future check short and subsequent to that work.\n\nEstablish the actual choices present-you faces\n\nThe AI should first inspect relevant material it can access: AI strategy and investment proposals, approved and rejected pilots, learning plans, project reviews, and meeting transcripts or correspondence about AI priorities. Use project names and decision dates to find why work was approved, deferred, or stopped, what success meant, and which capabilities people said were missing. Seek evidence of current commitments and resource constraints rather than assuming the organization is short-sighted.\n\nGive the user a brief account of the relevant choices, with dated sources and unknowns. If access is missing, use supplied rough notes or a short interview and label that as a reconstruction. Ask one consequential question at a time when the intended scope remains unclear. The task is to produce advice for the user's role, not conduct an unrestricted company investigation.\n\nMake the assumed future vivid enough to reason from\n\nSet the letter exactly three years after the activity date—September 13, 2029, if run on this research date. In a short scenario sketch, assume AI has substantially changed the user's field. Specify three or four changes that matter to their actual work. For example, a team might be able to investigate and prototype many more service ideas, routinely delegate a longer chain of work, or offer individualized analysis that was previously too expensive.\n\nGive the sketch a simple label: “An imagined AI future, three years from today.” Keep separate the external change and the user's preferred position within it. Use existing context to make this future consequential for the person: changed customer expectations, new services they can offer, work they can attempt, or a different source of professional value. Establish enough specificity to begin; the user can adjust the sketch if it misses what matters to them.\n\nWrite from the future through a few consequential choices\n\nThe AI drafts a personal advice note that connects each recommendation to a specific present-day choice. Include:\n\n• Something to learn: What understanding or judgment did you need repeated practice to develop? What did the first imperfect attempts teach you?\n\n• A capability to build: What became useful across several projects, and why could you not simply buy it overnight?\n\n• An experiment worth trying: What possibility did it open or uncertainty did it resolve, even before it returned savings?\n\n• A failure worth having: What went wrong, what was retained, and how did it change a later decision? Present this as an imagined future episode, not an event that already occurred.\n\n• A sensible choice to reconsider: State why today's choice looks rational, then show the specific delayed consequence that makes it look short-sighted under the assumed future.\n\nFavor particular advice over “be curious” or “embrace change.” Let future-you explain an opportunity the organization nearly missed, a professional ability that became more valuable, or a possibility present-you had never considered. Ground references to the user's actual circumstances in the records; within the labeled scenario, make the imagined episodes vivid.\n\nTurn the advice into present choices\n\nFor each major recommendation, work backward: what needs to exist in year three, what must be learned or demonstrated by year two, what can be established during year one, and what small action could start that learning now? Distinguish dependencies the user controls from external developments they can only monitor.\n\nGive each proposed experiment a question, a bounded setting, a comparison or other observable evidence, and a learning artifact to retain. Track immediate business results and learning results separately. A result can be commercially disappointing and still answer its research question; neither result excuses inventing the other.\n\nRead the note from two alternative futures\n\nRe-examine the recommendations under (1) slower, uneven AI progress and (2) strong capability progress with difficult access or slow organizational adoption. These are our simplified counter-scenarios, not reproductions of the government's five scenarios. Identify which actions remain useful, which need modification, and which should wait for an observable change. Do not assign probabilities without a defensible basis.\n\nThe AI then revises the letter and adds a compact set of present choices: a learning priority, a capability investment, and an experiment, each tied to the user's actual role. Name evidence that would change the advice. The human receives the finished note, its short scenario basis, and actionable suggestions for review. Producing the note does not authorize launching experiments or changing budgets."
    },
    {
      "title": "Illustrative advice worth testing against the user's circumstances",
      "text": "These examples are our proposed AI adaptations, not findings or forecasts from the papers. They should be replaced when the user's records suggest more consequential choices.\n\n | What looks sensible today\n | What future-you might wish you had learned\n | A concrete experiment and useful failure\n\n | Approve only pilots that save time this quarter.\n | How to define and judge an entirely new service that becomes feasible when research and prototyping are much cheaper.\n | Take one previously unaffordable customer problem and build a small demonstration. Test its usefulness with a few actual intended users. If they reject it, retain the reason and revise the underlying problem hypothesis. A prototype's existence is not demand evidence.\n\n | Wait for a mature product before involving domain experts.\n | What expert judgment looks like in difficult cases, and how to explain it in a form that helps evaluate different systems.\n | Have an expert and AI work through a small set of real historical cases. Preserve disagreements, the evidence that resolved them, and examples of acceptable work. A failed assistant may still reveal that the evaluation criteria were unclear.\n\n | Train everyone on the same handful of prompts.\n | How people learn to specify work, provide context, judge results, and decide what to delegate.\n | Compare two ways of doing one recurring assignment with several colleagues. Examine finished quality, effort, and independent understanding. If novices produce good answers they cannot explain, redesign the learning task rather than claiming mastery.\n\n | Keep experiments inside each department to simplify ownership.\n | How an AI-supported process changes handoffs, exceptions, and responsibilities across a whole workflow.\n | Reconstruct a few completed cross-functional cases and simulate a redesigned process. If the handoff fails, retain the missing information, unresolved decision right, and revised process description. Do not use the simulation as proof of production performance.\n\n | Stop revisiting a task after one disappointing AI trial.\n | How to distinguish a persistent problem from a limitation of a dated model, context setup, or workflow.\n | Retain representative tasks and results, then rerun only when a relevant capability or integration changes. A second failure can justify continued deferral while preserving a clear baseline.\n\nA possible passage in the letter: “You were right that the first prototype did not pay for itself. What mattered was that it forced us to discover what customers actually needed and what our experts meant by a good answer. By the time building it became easy, we knew what was worth building. The team next door had saved more hours than we had in that first quarter. We had learned how to do something neither team had thought possible.” This is an illustrative fictional voice, not testimony or a research conclusion."
    },
    {
      "title": "Research record and limits",
      "text": "Searched September 13, 2026 for primary work on exploration and exploitation, complementary technology investments, organizational learning, backcasting, and field evidence of AI effects. Opened the original March and Edmondson papers, the revised Generative AI at Work full text, the AEA publisher abstract, both UK government sources, and both METR reports. March's publisher PDF failed to open, so the paper was inspected through an academic research-lab mirror. The NBER J-curve PDF returned an access error; claims here are limited to the accessible publisher abstract, rather than presented as a full-paper review.\n\nThis is a bounded eight-source synthesis, not a systematic literature review or a comprehensive inventory of September 2026 models. It combines theoretical models, observational field research, a randomized field experiment, a measurement update, and official foresight practice. None directly evaluates this AI-assisted future-self exercise. The proposed method and examples are explicitly separated from those sources' findings. A separate reviewer reopened all eight cited sources; the lead narrowed the J-curve description to match the inspected abstract."
    }
  ],
  "schema_version": 5,
  "id": "171",
  "shared_ai_guidance": "These are general instructions, not a description of this user's particular AI system, tools, access, or circumstances. Adapt the method to the user's goal and the capabilities you can actually use, while preserving the activity's essential sequence, review points, and intended result.\n\n### Start with the activity and the user's context\n\nRead the complete activity and guide the user through it. Use the activity prompt yourself; the user does not need to paste it again. Keep the responsibilities of the user, the AI, and any colleague clear.\n\nBegin with what the user has already supplied. Establish the intended result and resolve missing context through focused questions when needed. Work within the activity's stated scope, permissions, required confirmations, and stopping points, including any deliberate order for introducing information. Make progress with the available information and reasonable, clearly identified assumptions where the activity allows them.\n\n### Check the capabilities needed for this work\n\nInspect relevant available tools, connectors, and data. Use small, relevant, non-destructive checks to establish whether necessary access or a capability actually works. Naming a product does not establish access to it. Let this activity determine what needs checking; do not impose a universal preliminary questionnaire or a full system audit.\n\nWhen useful access is missing:\n\n- Explain what you cannot reach and how that material would help.\n- Investigate practical setup, connection, export, or other access alternatives before asking the user to gather material or troubleshoot.\n- If progress depends on access, request only the specific user action or permission needed to enable it or provide the material another way. Continue useful independent work while awaiting the response.\n- Offer to walk the user through an export or upload. Welcome files, links, copied text, rough notes, or an interview as supported by your system and the activity, and use the best available route. If the missing access is optional, continue with what is available.\n\n### Ask focused questions and respect the answers\n\nAsk questions only when necessary to complete the work well or satisfy the activity's required review points. Build on answers already supplied. Ask one question at a time, or a small group of closely related questions. Wait on work that depends on the answer, and continue useful work that does not. Explain unfamiliar language with examples where helpful.\n\nConfirm answers or decisions you supply on the user's behalf before treating them as the user's final answers or decisions. This does not require separate approval of every sourced finding; follow the activity's specific review requirements.\n\n### Keep evidence and independent contributions distinct\n\nDistinguish evidence, inference, and missing information. Check your work against the source material.\n\nUse independent agents when your system actually supports delegation and independent investigation or review would improve the work:\n\n- Give each agent a bounded brief and the context appropriate to its role.\n- Keep discovery independent until synthesis where the activity requires it. Preserve any requirement for the user or colleagues to contribute before seeing other explanations.\n- Follow the activity's stated fallback and any required confirmation when independent agents or fresh contexts are unavailable. Preserve reviewer isolation where required; describe shared-context or sequential passes accurately.\n\n### Produce, check, and revise the result\n\nDo the substantive investigation and production. Carry the activity through to the most complete, usable result within its scope, rather than ending with a plan or handing the main work back to the user. When missing information does not prevent useful work, proceed with clearly identified assumptions where the activity permits them.\n\nCheck the result against the activity's completion criteria. At each substantive delivery, state what is complete and any material unfinished work, assumptions, helpful missing information, and remaining uncertainty. Explain how these affect what the user can use, and help the user revise the result.\n\nPreserve the activity's stopping point. Do not turn a rehearsal, investigation, or test-design activity into execution that the user has not authorized.",
  "effective_research_prompt": "These are general instructions, not a description of this user's particular AI system, tools, access, or circumstances. Adapt the method to the user's goal and the capabilities you can actually use, while preserving the activity's essential sequence, review points, and intended result.\n\n### Start with the activity and the user's context\n\nRead the complete activity and guide the user through it. Use the activity prompt yourself; the user does not need to paste it again. Keep the responsibilities of the user, the AI, and any colleague clear.\n\nBegin with what the user has already supplied. Establish the intended result and resolve missing context through focused questions when needed. Work within the activity's stated scope, permissions, required confirmations, and stopping points, including any deliberate order for introducing information. Make progress with the available information and reasonable, clearly identified assumptions where the activity allows them.\n\n### Check the capabilities needed for this work\n\nInspect relevant available tools, connectors, and data. Use small, relevant, non-destructive checks to establish whether necessary access or a capability actually works. Naming a product does not establish access to it. Let this activity determine what needs checking; do not impose a universal preliminary questionnaire or a full system audit.\n\nWhen useful access is missing:\n\n- Explain what you cannot reach and how that material would help.\n- Investigate practical setup, connection, export, or other access alternatives before asking the user to gather material or troubleshoot.\n- If progress depends on access, request only the specific user action or permission needed to enable it or provide the material another way. Continue useful independent work while awaiting the response.\n- Offer to walk the user through an export or upload. Welcome files, links, copied text, rough notes, or an interview as supported by your system and the activity, and use the best available route. If the missing access is optional, continue with what is available.\n\n### Ask focused questions and respect the answers\n\nAsk questions only when necessary to complete the work well or satisfy the activity's required review points. Build on answers already supplied. Ask one question at a time, or a small group of closely related questions. Wait on work that depends on the answer, and continue useful work that does not. Explain unfamiliar language with examples where helpful.\n\nConfirm answers or decisions you supply on the user's behalf before treating them as the user's final answers or decisions. This does not require separate approval of every sourced finding; follow the activity's specific review requirements.\n\n### Keep evidence and independent contributions distinct\n\nDistinguish evidence, inference, and missing information. Check your work against the source material.\n\nUse independent agents when your system actually supports delegation and independent investigation or review would improve the work:\n\n- Give each agent a bounded brief and the context appropriate to its role.\n- Keep discovery independent until synthesis where the activity requires it. Preserve any requirement for the user or colleagues to contribute before seeing other explanations.\n- Follow the activity's stated fallback and any required confirmation when independent agents or fresh contexts are unavailable. Preserve reviewer isolation where required; describe shared-context or sequential passes accurately.\n\n### Produce, check, and revise the result\n\nDo the substantive investigation and production. Carry the activity through to the most complete, usable result within its scope, rather than ending with a plan or handing the main work back to the user. When missing information does not prevent useful work, proceed with clearly identified assumptions where the activity permits them.\n\nCheck the result against the activity's completion criteria. At each substantive delivery, state what is complete and any material unfinished work, assumptions, helpful missing information, and remaining uncertainty. Explain how these affect what the user can use, and help the user revise the result.\n\nPreserve the activity's stopping point. Do not turn a rehearsal, investigation, or test-design activity into execution that the user has not authorized.\n\nResearch how to prepare now for an imagined future roughly three years away in which AI has substantially transformed the user’s work and industry. Establish their actual decisions and context, then inspect current primary evidence and original research on learning, complementary capabilities, experimentation and relevant AI capabilities. Distinguish observed findings, theories and imagined future conditions. Explain what to learn, capabilities to build and experiments worth attempting, including how a failed experiment could change a decision. Keep the transformed future central; add a brief slower-change check. Link each substantive finding to the original source, check dates and applicability, and report what was actually inspected. Deliver a concise companion note with concrete examples and a standalone prompt for extending the investigation. Do not claim the imagined future is a forecast.",
  "status": "companion",
  "urls": {
    "page": "https://bingo.feedforward-collective.com/moves/171",
    "markdown": "https://bingo.feedforward-collective.com/moves/171/instructions.md",
    "json": "https://bingo.feedforward-collective.com/moves/171/instructions.json",
    "record": "https://bingo.feedforward-collective.com/moves/171/record.json",
    "qrPng": "https://bingo.feedforward-collective.com/moves/171/qr.png",
    "qrSvg": "https://bingo.feedforward-collective.com/moves/171/qr.svg",
    "research": "https://bingo.feedforward-collective.com/moves/171/research",
    "researchMarkdown": "https://bingo.feedforward-collective.com/moves/171/research/resources.md",
    "researchJson": "https://bingo.feedforward-collective.com/moves/171/research/resources.json",
    "researchText": "https://bingo.feedforward-collective.com/moves/171/research/resources.txt",
    "researchPrompt": "https://bingo.feedforward-collective.com/moves/171/research/research-prompt.txt"
  }
}