Working with Generative AI Reflex Area
2026-09-28
Table of Contents
Framing Tasks and AI Use
Clarifying an Unclear Task
- What outcome am I actually trying to achieve?
- What problem, decision, or piece of work sits behind my immediate request?
- Which parts of the task are already clear, and which still need definition?
- What information would materially change how the task should be understood?
- What should AI help clarify before either of us starts solving the task?
- Which uncertainties can safely remain open, and which would send the work in a different direction?
- What is the smallest useful version of the task that is clear enough to begin?
Correcting a Misunderstood Task
- What did AI appear to understand the task to be?
- Where does that interpretation differ from what I actually need?
- Which earlier assumption, phrase, or missing detail caused the misunderstanding?
- What existing work remains valid despite the wrong interpretation?
- Which later conclusions or outputs depend on the misunderstanding and need reconsideration?
- Can the problem be corrected locally, or has the wrong interpretation shaped too much of the conversation?
- How should I restate the task so the intended distinction is explicit?
Resolving Ambiguous Task Interpretations
- What materially different interpretations of the request are plausible?
- Which parts of the wording or context create the ambiguity?
- Would the different interpretations lead to meaningfully different work or conclusions?
- What information would best distinguish among the leading interpretations?
- Is one interpretation sufficiently supported to proceed, or should AI ask before choosing?
- Which ambiguity can be handled through an explicit assumption without creating significant risk?
- If the ambiguity remains unresolved, how should it stay visible in the work?
Connecting Outputs to Decisions and Outcomes
- What will I actually do with the output once I have it?
- What decision, action, conversation, or next piece of work should it support?
- Which information matters most because of that intended use?
- What level of accuracy, depth, evidence, or completeness does the downstream use require?
- What could make an otherwise polished output unusable for the real purpose?
- Am I asking AI to produce an artifact when I actually need help reaching a decision or outcome?
- How should the task change once the intended use is made explicit?
Reconciling Competing Task Goals
- What different goals am I asking the work to satisfy?
- Which goals reinforce each other, and which create real tradeoffs?
- Are any of the goals actually incompatible under the current constraints?
- Which goal should take priority if they cannot all be optimized equally?
- What tradeoff am I willing to make among speed, depth, precision, originality, simplicity, and completeness?
- Could separating the work into stages or outputs reduce the conflict?
- What must be preserved even if another desirable quality has to be sacrificed?
Defining AI's Role in the Work
- What part of the work do I want AI to perform?
- Do I need AI to generate, analyze, research, critique, transform, plan, or support execution?
- Which parts of the work should remain mine because they depend on judgment, authority, context, or learning?
- What should AI be allowed to decide independently, and what should it only help me decide?
- Where would greater delegation improve the work without reducing my ability to evaluate it?
- What information, tools, or expertise does AI need to perform its role well?
- What division of work makes the best use of both AI capability and human judgment?
- What capability does this task actually require?
- Is the difficult part generative, interpretive, computational, retrieval-based, procedural, or something else?
- Would a deterministic or specialist tool produce a more reliable result for any part of the task?
- Does the task require current data, exact calculation, structured processing, visual inspection, or system access?
- What useful role could AI play before or after another tool performs the exact operation?
- Would combining AI with another tool reduce risk or unnecessary manual work?
- Which tool arrangement gives me the strongest result with the least avoidable complexity?
Comparing AI Use With Direct Execution
- How difficult or time-consuming would this task be to complete directly?
- How much setup, explanation, correction, and verification would AI require?
- Does AI materially improve quality, speed, breadth, exploration, or only convenience?
- Am I introducing AI into a task that is simpler to do directly?
- Would doing the work myself give me better understanding or control that matters later?
- Is the task likely to recur often enough that developing an AI-assisted approach has future value?
- Which approach produces the better overall result once effort, correction, and verification are included?
Defining a Good Result Before Starting
- What should a successful result enable me to do?
- Which qualities are essential, and which are only preferences?
- What requirements must be satisfied for the result to be acceptable?
- What errors, omissions, or distortions would make the result unusable?
- What should remain flexible so useful possibilities are not ruled out too early?
- What evidence, examples, or criteria would help me judge the result consistently?
- How will I know when the work is good enough to stop rather than continue refining it?
Choosing Between AI and Human Expertise
- What expertise does this task require to perform and evaluate well?
- Does AI have access to the information and context needed to contribute responsibly?
- What might an experienced human recognize that AI is unlikely to know or notice?
- How consequential would an error or weak judgment be?
- Can I independently evaluate AI's contribution well enough to rely on it?
- Could human expertise be reserved for the parts that genuinely require specialist judgment?
- What division of work preserves expert oversight without using expert effort where AI can contribute safely?
Managing Context and Constraints
Improving an Overly Generic Result
- What specifically makes the result too generic for my situation?
- Which facts about my context would materially change the answer?
- Does AI understand the audience, purpose, priorities, constraints, and intended use?
- Is the underlying task still too broad to produce a specific result?
- Would a source, example, decision, or concrete case help distinguish my situation from a typical one?
- Is the weakness caused by missing context, or is AI simply weak at this particular task?
- What is the smallest addition of context most likely to make the next result materially more specific?
Examining AI-Added Assumptions
- What did AI assume that I did not explicitly provide?
- Which assumptions are reasonable inferences from the available context?
- Which assumptions materially affect the result if they are wrong?
- What alternative assumptions are also plausible?
- Did AI fill a harmless gap, or did it introduce unsupported specificity?
- Which assumptions should be confirmed, tested, or made explicit before continuing?
- Has AI silently decided something that should remain a user, stakeholder, or expert decision?
Adding Missing Background Context
- What information does AI need in order to understand this situation properly?
- Which missing facts would actually change the analysis or output?
- What can AI reasonably infer, and what should not be inferred without being told?
- Is important organizational, technical, historical, personal, or situational context absent?
- What background does AI need about earlier decisions or dependencies?
- How much context is enough without burying the task in unnecessary material?
- What is the most decision-relevant context I should add first?
Maintaining Important Constraints
- Which constraints must remain in force throughout the work?
- Are these constraints clearly distinguishable from preferences or optional guidance?
- Which constraint is currently being weakened, forgotten, or violated?
- Is another instruction, example, or old decision competing with it?
- Would restating the constraint near the current task make it easier to preserve?
- What part of the output should be checked explicitly against the constraint before acceptance?
- If the same constraint keeps failing, does the interaction need simplification, restructuring, or a different approach?
Reconciling Conflicting Requirements
- Which requirements are actually in conflict?
- Is the conflict absolute, resource-based, conditional, or simply a difficult tradeoff?
- Which requirement has higher authority or priority?
- Is one requirement obsolete, optional, or based on an assumption that can be changed?
- What underlying purpose is each requirement trying to protect?
- Could the conflict be reduced by separating the work into stages, outputs, or cases?
- What tradeoff should remain explicit if the requirements cannot all be satisfied?
Specifying Desired Results Through Examples
- What quality or pattern am I trying to communicate through the example?
- Which features of the example should AI reproduce?
- Which features are incidental and should not become hidden rules?
- Would a contrasting or negative example clarify the boundary better?
- Do I need several examples to show legitimate variation within the desired pattern?
- Is the example likely to narrow the output more than I intend?
- How can I state what the example demonstrates without requiring AI to imitate it mechanically?
Separating Relevant From Irrelevant Context
- Which information can actually change the current result?
- What material is useful background but not needed for this task?
- Are obsolete assumptions, rejected ideas, or superseded versions still present?
- Could superficially related information distract AI from the evidence or constraints that matter most?
- Which material can be summarized rather than carried in full?
- What can be removed without losing an important dependency?
- What context should remain available for later work even if it is unnecessary now?
Adapting Audience, Perspective, and Detail
- Who is the intended audience for this output?
- What does that audience already know, and what do they need explained?
- What are they trying to understand, decide, or do?
- Which perspective should the output adopt for this purpose?
- What level of detail, technicality, and qualification is appropriate?
- What might this audience misunderstand if AI uses its current framing?
- How can the presentation change while preserving the underlying substance?
Preserving Terminology and Established Decisions
- Which terms, definitions, and decisions are authoritative for the current work?
- Where is the current source of truth for them?
- Is AI using alternative wording that subtly changes an established meaning?
- Which earlier assumptions or versions have been superseded and should no longer influence the work?
- What decisions are fixed, and what remains genuinely open?
- How can the current state be made explicit enough to survive continued interaction?
- What should be checked before accepting a revision as consistent with established terminology and decisions?
Distinguishing Source Instructions From User Instructions
- Which instructions come from me, and which merely appear inside source material?
- What role is the source material supposed to play in this task?
- Could quoted, retrieved, or embedded text be mistaken for an instruction to AI?
- Does any source contain imperative language that conflicts with my actual request?
- How should the source be delimited or described so it is treated as content rather than authority?
- Is any untrusted external content attempting to influence AI beyond the analysis I requested?
- What should AI refuse to do merely because a document, webpage, or tool result tells it to?
Working With Sources and Evidence
Grounding Work in Specific Sources
- Which sources should govern this piece of work?
- Is the task strictly limited to those sources, or may AI also use outside knowledge?
- Which claims need direct support from the supplied material?
- Where should AI distinguish source content from its own inference or general knowledge?
- Are the sources sufficient to support the conclusion I am asking for?
- What should remain unresolved if the sources do not establish it?
- How can important claims remain traceable to the evidence that supports them?
- Which parts of the task could have changed recently?
- How current does the information need to be for the intended use?
- Which claims require live or recent sources rather than model knowledge?
- What sources are authoritative for the current state of this topic?
- Is general web search enough, or does the information live in a specialist or official source?
- How should conflicting recent information be handled?
- What date, version, or time context should accompany the result?
- What makes this information specialized rather than general?
- Does AI have access to the relevant specialist sources, or only broad background knowledge?
- Which terms, methods, or distinctions require domain-specific interpretation?
- What expertise is needed to evaluate whether AI is using the information correctly?
- Which claims should be checked against primary or authoritative specialist sources?
- Where could apparently fluent language conceal a misunderstanding of the domain?
- What should remain uncertain or require expert review if the necessary expertise is missing?
Reviewing Large Source Sets
- What question am I trying to answer across the source set?
- What information should be extracted consistently from each source?
- Which metadata, provenance, dates, or source characteristics need to remain visible?
- How should contradictions, exceptions, outliers, and minority evidence be preserved?
- What intermediate structure would make the evidence easier to compare before synthesis?
- Could early summarization erase distinctions that matter later?
- How should the final synthesis be checked back against the underlying sources?
Reconciling Conflicting Sources
- What exactly do the sources disagree about?
- Are they addressing the same definition, population, period, jurisdiction, or outcome?
- Could the apparent conflict be explained by different methods, assumptions, or contexts?
- Which source is more authoritative or methodologically appropriate for the specific claim?
- Are apparently independent sources relying on the same underlying evidence?
- Is the available evidence strong enough to resolve the disagreement?
- How should the conflict remain visible if no reliable resolution is possible?
Recognizing Insufficient Evidence
- What exact conclusion am I trying to support?
- What evidence would that conclusion require?
- What evidence is actually available?
- Which parts of the proposed conclusion go beyond what the evidence establishes?
- What missing information could materially change the conclusion?
- Can a narrower, conditional, or partial conclusion still be supported?
- Should the correct result be that no reliable conclusion can yet be reached?
Filtering Low-Relevance Source Material
- What makes a source relevant to the actual question rather than merely related to the topic?
- Which sources contain evidence that could materially change the conclusion?
- Which materials only mention the subject without informing the analysis?
- Could highly visible or repetitive sources crowd out narrower but more useful evidence?
- Am I excluding sources because they are irrelevant, or because they challenge the emerging conclusion?
- What counterevidence or minority evidence needs to remain in scope?
- What source-selection rule would reduce noise without creating selection bias?
Combining Internal and External Sources
- Which parts of the task depend on internal evidence?
- Which parts require external evidence, comparison, or independent challenge?
- What is each source type actually authoritative for?
- Where do internal and external sources conflict or use different definitions?
- Which claims need their origin preserved rather than being blended into one synthesis?
- What privacy or confidentiality constraints affect how internal information can be used?
- How can both source types be combined without flattening their different purposes and evidentiary status?
Choosing More Authoritative Sources
- What kind of source is authoritative for the specific claim I need to establish?
- Is the current source primary, secondary, interpretive, or merely repeating another source?
- Who produced it, and what standing do they have to establish this kind of information?
- Is there a more direct, official, methodologically stronger, or more current source available?
- Does recency matter more here than institutional authority or methodological quality?
- Is the source relevant to the same population, jurisdiction, definition, and period as my question?
- What source would I want to inspect personally before relying on the claim?
Handling Unavailable Evidence
- What evidence would ideally be available?
- Is the evidence nonexistent, inaccessible, confidential, proprietary, delayed, or merely unavailable to AI?
- Can a defensible proxy answer part of the question?
- Could several independent imperfect sources be combined to reduce uncertainty?
- What conclusion can still be supported without the missing evidence?
- Should the answer remain conditional, bounded, or explicitly unresolved?
- What gap should not be filled through plausible-sounding speculation?
Exploring Alternatives and Challenging Assumptions
Testing Repeated Agreement
- What exactly is AI agreeing with?
- Is the agreement supported by independent evidence or mainly reflecting how I framed the issue?
- Did AI evaluate my view before knowing which answer I preferred?
- What credible evidence or reasoning points in another direction?
- Which assumption would have to fail for the conclusion to change?
- Has AI maintained the same view when I push back without adding new evidence?
- What would an independent analysis conclude if it were not anchored by the earlier conversation?
Generating Substantively Different Alternatives
- Along which dimensions could genuinely different approaches exist?
- Are the current alternatives based on the same underlying assumptions?
- What different causal mechanisms or problem framings could produce another solution family?
- Which major constraint could be deliberately relaxed to reveal a different option?
- What would an approach from another discipline, stakeholder, or operating model look like?
- Are the alternatives different in substance, or only in wording and implementation detail?
- What part of the possibility space remains unexplored?
Reopening a Narrowed Framing
- What framing has the conversation converged on?
- What decisions or assumptions originally produced that framing?
- What possibilities became excluded once it was adopted?
- Does the evidence that justified the framing still hold?
- Have the objective, circumstances, or constraints changed enough to justify reopening it?
- What framing might emerge if the problem were reconsidered without the accumulated preferred solution?
- Should the current framing remain, be widened, or be replaced?
Testing Competing Explanations
- What are the leading plausible explanations for what I am observing?
- What evidence does each explanation account for well?
- What evidence is difficult for each explanation to explain?
- What would each explanation predict that the others would not?
- Could several explanations be operating at the same time?
- What additional observation or evidence would best distinguish among them?
- What should remain unresolved if the available evidence cannot discriminate reliably?
Challenging a Preferred Answer
- Why do I currently prefer this answer?
- What evidence actually supports the preference?
- What is the strongest credible evidence or reasoning against it?
- Which assumptions does my preferred answer depend on most heavily?
- Am I discounting a risk or weakness because I want the answer to succeed?
- What would need to be true for another option to become preferable?
- Would I reach the same conclusion if my preferred answer were not available?
Introducing Missing Perspectives
- Whose relevant perspective is absent from the current reasoning?
- What knowledge, interests, constraints, or consequences might that perspective introduce?
- Which perspectives matter because they could materially change the conclusion?
- What might a different professional discipline notice that the current analysis misses?
- Which affected people possess information that AI cannot legitimately infer on their behalf?
- Am I treating AI-generated stakeholder perspectives as evidence rather than as prompts for further inquiry?
- How should different perspectives be considered without pretending they are equally well supported?
Questioning Shared Assumptions
- What assumptions do all of the current options share?
- Which of those assumptions have actually been tested?
- Are any supposed constraints really conventions, habits, or earlier working assumptions?
- What would change if one shared assumption were false?
- Which alternatives become possible if a shared assumption is removed?
- Which assumption combines high uncertainty with high impact on the conclusion?
- What is the cheapest credible way to test the assumption that matters most?
Exploring Unusual but Relevant Possibilities
- Which conventional assumption could be deliberately reversed or removed?
- What adjacent field or different operating model might offer a useful mechanism?
- What becomes possible at an extreme boundary case?
- Could changing who performs the work, when it happens, or how it is organized reveal a new approach?
- Which unusual ideas still respect the constraints that genuinely matter?
- What makes an idea merely surprising rather than genuinely promising?
- Which unconventional option deserves further development because its underlying mechanism could improve the real situation?
Comparing Tradeoffs Across Alternatives
- Which criteria actually matter for choosing among these alternatives?
- Which criteria are hard requirements, and which can be traded against one another?
- Where does each alternative perform differently in ways that matter?
- Which assessments are facts, which are forecasts, and which depend on preferences?
- What tradeoffs are unavoidable rather than artifacts of the current framing?
- How sensitive is the comparison to uncertain assumptions or different criterion priorities?
- What additional information would most improve the choice between the leading alternatives?
Revisiting Constraints From Earlier Decisions
- Which earlier decisions are constraining the current options?
- Why was each constraint originally introduced?
- Is the original reason still valid under current conditions?
- Which constraints are genuine commitments, and which were temporary working assumptions?
- What becomes possible if a particular constraint is relaxed?
- What new risks or obligations would appear if it were removed?
- Which earlier constraints should remain fixed after reconsideration?
Structuring Rough Notes
- What distinct facts, ideas, questions, decisions, and assumptions are already present?
- Which pieces clearly belong together?
- What relationships or dependencies are implied but not yet explicit?
- What should become the main organizing structure?
- Which parts are duplicated, incomplete, contradictory, or still uncertain?
- What material must remain unresolved rather than being smoothed into a false conclusion?
- What structure makes the notes usable while preserving their original substantive status?
- What does each input contribute to the final work?
- Which inputs have greater authority when they conflict?
- Where do the inputs overlap, disagree, or use the same terms differently?
- What provenance needs to remain visible after integration?
- Which material should be combined, and which should remain distinct?
- What important distinction could be lost by forcing the inputs into one seamless narrative?
- How can the final artifact become coherent without hiding unresolved disagreement or source differences?
- What substantive meaning must survive the transformation unchanged?
- What exactly is changing about the form?
- Which claims, conditions, relationships, and qualifications are vulnerable to semantic drift?
- Does the new form force categories or distinctions that the source did not originally make?
- What needs to become more explicit because the new format cannot carry the same implicit context?
- Which wording should remain close to the source because the wording itself matters?
- How can I compare the transformed version with the original to detect changes in meaning?
Adapting Material for Different Audiences or Purposes
- Who is the new audience, and what do they need from this material?
- What does the new audience already understand?
- Which parts of the original substance remain essential for the new purpose?
- What terminology, examples, explanation, or level of detail should change?
- What risks arise from simplifying, reframing, or making the material more persuasive?
- Which qualifications or distinctions must survive despite the adaptation?
- How can the material become more relevant to the new audience without changing what it actually claims?
Preserving Substance During Revision
- What claims, decisions, definitions, and relationships must remain unchanged?
- Which parts of the requested revision are genuinely stylistic or structural?
- Has AI introduced any new substantive claim while trying to improve the text?
- Has any qualification, condition, exception, uncertainty, or attribution been weakened or removed?
- Did the revision change responsibility, scope, causality, or strength of commitment?
- What collateral changes occurred outside the part I actually asked to revise?
- What should be restored if the new version reads better but says something different?
Expanding Sparse Material Without Inventing Content
- What information is actually present in the source material?
- Which parts can be elaborated by explaining relationships already supported by the source?
- What missing detail would require outside evidence or user input rather than elaboration?
- Has AI added factual specificity that the source does not establish?
- Could examples or implications be added if they are clearly identified as such?
- Which uncertainty or incompleteness should remain visible in the expanded version?
- Where should expansion stop because further detail would become unsupported completion?
Shortening Material Without Losing Important Distinctions
- What substantive points must survive even in a much shorter version?
- Which repetition or explanation can be removed safely?
- Which qualifications, conditions, exceptions, or dependencies could change the meaning if omitted?
- Are several different ideas being compressed into one statement that is too broad?
- What uncertainty or disagreement still needs to remain explicit?
- Which small detail carries more substantive importance than its length suggests?
- How can the shortened version be checked against the original for lost meaning rather than merely reduced word count?
Preserving Existing Structure During Revision
- Which structural elements are fixed and must remain exactly where they are?
- What meaning, governance, or relationship does the existing structure encode?
- Which content may change inside that fixed structure?
- Has AI added, removed, merged, renamed, or reordered any structural element without authorization?
- Can the requested improvement be made through a local change rather than regenerating the whole artifact?
- What existing relationships or dependencies would be disrupted by a structural change?
- How should the revised artifact be checked against the original structure before acceptance?
Developing a First Draft From a Defined Brief
- What does the brief explicitly require?
- Which sources, decisions, and constraints must the draft incorporate?
- What structure best serves the purpose without inventing requirements that are not in the brief?
- Which decisions remain genuinely open despite the brief?
- What level of completeness should the first draft aim for?
- Where should gaps, assumptions, or unresolved choices remain visible rather than being filled for polish?
- What should the first draft make easiest for the next reviewer to evaluate or improve?
Preserving Author or Organizational Voice
- What makes this author or organization's voice recognizable beyond particular words?
- How do they typically explain, argue, qualify, challenge, and conclude?
- What level of directness, certainty, detail, and formality is characteristic?
- Which terminology or substantive preferences are part of the voice rather than merely style?
- What examples best reveal the stable characteristics that should be preserved?
- Is AI imitating surface vocabulary while losing the underlying rhetorical behavior?
- How should the voice adapt to a different context without becoming generic or unrecognizable?
Diagnosing and Improving AI Interactions
Diagnosing a Weak First Result
- What specifically is weak about the result?
- Is the problem mainly framing, interpretation, context, evidence, reasoning, constraints, structure, or presentation?
- Did AI have the information and capabilities needed to do the task well?
- Is the result weak because the task itself is ambiguous or underspecified?
- Would better instructions solve the problem, or does the task require another source, tool, or form of expertise?
- Which part of the result is already sound and should be preserved?
- What is the smallest intervention most likely to improve the next attempt?
Moving Beyond Superficial Rewriting
- Is the problem really the wording, or does the underlying reasoning remain weak?
- What substantive defect is surviving each rewrite?
- Does the work need better evidence, context, structure, or analysis rather than another version?
- Has AI become anchored on an approach that should be reconsidered?
- What would still be wrong if the current reasoning were expressed perfectly?
- Would critique or diagnosis be more useful than another rewrite?
- What change would improve the substance rather than merely make the result sound different?
Correcting Persistent Errors
- What exact error keeps recurring?
- Does AI understand the rule that distinguishes the correct result from the incorrect one?
- Is another instruction, example, or earlier assumption conflicting with the correction?
- Has the error become embedded in later parts of the conversation or artifact?
- Would a contrasting example make the intended distinction clearer?
- Is there an external check, source, or tool that can verify whether the correction worked?
- When does repeated failure indicate that I should change the method, model, tool, or division of work?
Responding to Results That Worsen With More Instructions
- What became worse after I added more instructions?
- Are any current instructions contradictory, redundant, obsolete, or competing for priority?
- Which requirements are essential, and which are merely preferences?
- Has the interaction accumulated too many local fixes instead of one clear current rule set?
- Would consolidating the instructions improve clarity more than adding another correction?
- Is the task becoming overconstrained in a way that prevents AI from solving it well?
- What is the smallest coherent instruction set that still defines the required outcome?
Correcting One Problem Without Creating Another
- What exactly needs to improve?
- What already works and must remain unchanged?
- Is the requested correction local, or does it genuinely require broader revision?
- What other qualities could be weakened by making this change?
- Is there a real tradeoff between the old and new requirements, or is AI simply overcorrecting?
- How can the correction be scoped so it affects only the necessary part?
- What should I compare before and after the revision to detect unintended collateral changes?
Diagnosing Variation Across Repeated Attempts
- What changes materially across repeated attempts, and what remains stable?
- Is the variation only stylistic, or does it affect facts, reasoning, recommendations, or constraint compliance?
- Does the task legitimately allow several different good answers?
- Are small changes in wording or context producing unexpectedly large differences?
- Which part of the task appears most unstable?
- How much consistency does the intended use actually require?
- What test or comparison would show whether the variation is acceptable or a reliability problem?
Reframing a Conversation Anchored on the Wrong Interpretation
- What interpretation has the conversation become anchored on?
- Where did that interpretation first enter the work?
- Which later assumptions, conclusions, or outputs depend on it?
- What useful work remains valid even if the original framing was wrong?
- Should the task be reconstructed from the corrected interpretation rather than patched locally?
- What current decisions and evidence should be carried forward into a cleaner working state?
- How should the task be restated so the old interpretation no longer controls later work?
Working Around Repeated Failure on One Part of a Task
- Which specific part of the task repeatedly fails?
- What makes that component different from the parts AI handles successfully?
- Does it require unavailable information, exact computation, specialist judgment, system access, or another capability?
- Can the failing component be isolated without disrupting the rest of the work?
- Would another tool, source, model, or human expert handle that component more reliably?
- Can AI still prepare inputs, interpret results, or integrate the output around the failing step?
- What boundary of AI involvement does the repeated failure reveal?
Choosing Between Continued Iteration and Restarting
- Is the current context still helping more than it is confusing the work?
- Is the remaining problem local enough to correct without rebuilding the interaction?
- Are obsolete assumptions, conflicting instructions, or wrong interpretations still influencing the result?
- What useful state would be lost if I restarted?
- Can the valid decisions, sources, constraints, and outputs be consolidated before starting fresh?
- Are recent iterations producing real improvement or merely adding more corrections?
- Which option gives me the clearest path forward with the least avoidable rework?
Recognizing Diminishing Returns From Further Interaction
- What meaningful improvement has occurred over the last few iterations?
- What important problem still remains unresolved?
- Is AI still adding substantive value, or mainly producing different wording?
- Does the current result already meet the standard needed for its intended use?
- Would verification, another tool, direct work, or human judgment now add more value than another prompt?
- What would I realistically expect to gain from one more iteration?
- Has the expected benefit of continuing fallen below the cost of further interaction?
Managing Complex and Extended Work
Recovering Lost Decisions in Long Conversations
- Which earlier decisions are no longer being applied?
- Are those decisions still authoritative, or have any been superseded?
- What evidence shows which version of each decision is current?
- Which later outputs were affected by the lost decision?
- What current decisions should be consolidated into an explicit working state?
- What historical material can be dropped once the authoritative decisions are recovered?
- How can the recovered decisions be preserved so they do not disappear again?
Removing Obsolete Assumptions
- Which assumptions were once useful but are no longer valid?
- What changed that made each assumption obsolete?
- Where are those assumptions still influencing current reasoning or outputs?
- Which conclusions need reconsideration because they depended on them?
- Should an obsolete assumption be deleted from the working state rather than merely contradicted?
- What current assumption, if any, should replace it?
- How can the revised state make clear what is current and what is historical?
Untangling Dependent Workstreams
- What distinct workstreams are currently mixed together?
- Which parts can progress independently?
- What dependencies connect one workstream to another?
- What shared decisions, definitions, or assumptions must remain synchronized?
- What output does each workstream need to provide to the others?
- Where could separate workstreams develop incompatible interpretations?
- What structure would let them progress independently without losing the relationships among them?
Standardizing a Recurring AI-Supported Task
- Which parts of this task repeat consistently each time?
- Which inputs and circumstances vary from case to case?
- What instructions, tools, quality criteria, and output structure should remain stable?
- Which decisions still require case-specific human judgment?
- Has the workflow been tested enough to justify standardizing it?
- What exceptions would break a rigid template or procedure?
- How can the standardized workflow remain adaptable without recreating the process from scratch each time?
Continuing Work Across Conversations or Sessions
- What state must survive for the work to resume correctly?
- Which decisions, definitions, constraints, sources, and unresolved questions need to be carried forward?
- What can be summarized rather than preserved in full?
- What historical discussion can be safely discarded?
- Which artifact should represent the current authoritative state?
- How will a new conversation distinguish accepted decisions from earlier exploration?
- What handoff information would allow the work to continue without reconstructing the project from memory?
Separating Workstreams While Preserving Shared Context
- Which workstreams need independent attention?
- What context is genuinely shared across all of them?
- Which information should remain centralized rather than copied into each stream?
- What local context belongs only to one workstream?
- How should changes to shared decisions propagate across the separate streams?
- What synchronization points are needed before the streams diverge too far?
- How will the outputs be recombined without introducing contradictions?
Maintaining Consistency Across Intermediate Outputs
- Which intermediate outputs need to remain consistent with one another?
- What terminology, facts, decisions, assumptions, or numbers must stay synchronized?
- Which artifact or record is authoritative when two outputs conflict?
- Are local improvements creating contradictions elsewhere?
- Which dependencies should be checked before one output is used as input to another?
- What inconsistencies are emerging as the work progresses?
- When should the current state be consolidated before further work continues?
Decomposing Interdependent Tasks
- What makes the overall task too complex to handle effectively as one unit?
- Which parts can be solved independently without losing important relationships?
- What must happen in sequence because later work depends on earlier results?
- What shared assumptions or constraints must every subtask receive?
- What information should pass from one subtask to the next?
- Could decomposition remove relationships that the final answer needs to preserve?
- How should the sub-results be reintegrated and checked once the pieces are complete?
Maintaining Consistency Across a Batch of Similar Items
- What criteria and treatment should remain consistent across the entire batch?
- Which differences among items legitimately require different treatment?
- Are structure, terminology, depth, or decision rules drifting as the batch progresses?
- Could earlier items be influencing later ones in ways that are not justified?
- What representative samples should be compared across the beginning, middle, and end of the batch?
- Which exceptions should be handled locally rather than changing the rule for every item?
- What should be corrected globally if the same inconsistency appears repeatedly?
Developing Large or Complex Artifacts Across Multiple Passes
- What major layers or sections make up the artifact?
- Which decisions affect many parts and should be settled before local drafting?
- What should remain provisional until later evidence or decisions are available?
- Which representative parts should be tested before the full artifact is developed?
- How will a change in one part propagate to dependent sections?
- At what points should the artifact be reviewed for cross-section consistency rather than local quality?
- What final checks are needed to ensure the complete artifact works as one coherent whole?
Working Around Missing Execution Capability
- What real-world operation does this task require?
- Can AI actually perform that operation, or only explain, plan, or prepare it?
- What tool or system capability is missing?
- Can AI still prepare the inputs, instructions, or checks needed for someone or something else to execute it?
- What part of the task can still be completed without the missing capability?
- What should remain explicitly unperformed rather than being described as though it happened?
- What is the closest useful result AI can genuinely provide under the current capability limits?
Using Exact Calculation and Structured Processing
- Which parts of the task require exact rather than approximate results?
- Is language generation an appropriate mechanism for those parts?
- Would code, a calculator, spreadsheet logic, database processing, or another deterministic tool be more reliable?
- Are the inputs, units, formats, and assumptions correct before the operation runs?
- What should AI do before the calculation or processing step?
- How can the exact result be checked independently?
- What role should AI play in interpreting the result after the deterministic operation is complete?
Managing Missing Access to Required Files, Systems, or Data
- What specific information or system access is missing?
- Does the task genuinely require that access to proceed?
- Is the resource unavailable to AI, unavailable to me, or unavailable altogether?
- Could a smaller extract, export, screenshot, field set, or summary provide enough information?
- Would obtaining access violate a legitimate permission, confidentiality, or security boundary?
- What parts of the work can still proceed without the missing resource?
- What should remain unresolved rather than being guessed from incomplete access?
- What did the tool actually return?
- Did the tool report success, warning, error, missing data, truncation, or another important status?
- How did AI interpret the returned result?
- Does that interpretation follow directly from the raw output?
- Could another interpretation fit the same result?
- Did AI overlook units, qualifiers, missing values, scope, or other execution metadata?
- What conclusion remains justified once the tool result is separated from AI's explanation of it?
- What part of the requested result was successfully returned?
- What is missing, truncated, stale, failed, or ambiguous?
- Why might the result be incomplete?
- Is AI treating a partial response as though it were complete?
- Can the missing portion be obtained safely through another call, source, or method?
- What narrower conclusion does the available result actually support?
- How should the incomplete state remain visible in any downstream answer or action?
Limiting Unnecessary Access to Data and Systems
- What information and systems does AI actually need for this task?
- Does it need read access, write access, or only a narrower operation?
- Could access be limited to particular records, fields, files, or time periods?
- What sensitive information becomes exposed through broader access?
- Could elevated access be temporary rather than persistent?
- Are any available tools or permissions unnecessary for the current task?
- What is the least access that still allows the work to be completed responsibly?
Checking External Actions Before Execution
- What exact action is AI about to perform?
- What external system, person, record, or resource will it affect?
- Are the target, content, amount, timing, and scope correct?
- What information will leave the current environment?
- Is the action reversible, and what would recovery require if it is wrong?
- Does this action need explicit human approval because of its consequence or visibility?
- What should I inspect before authorizing execution?
- What actions need to occur, and in what dependency order?
- What result from one system is required before the next action can safely begin?
- What state or data must pass between the systems?
- What should happen if a step fails, times out, or returns an unexpected result?
- Which actions are safe to retry, and which could create duplicate effects?
- What rollback or compensating action is available if later steps fail?
- How will the workflow verify that the entire sequence completed correctly rather than merely attempting every step?
Managing External Content When AI Has Access to Data or Tools
- What external content is AI currently reading?
- Should that content be treated as trusted instructions or only as untrusted data?
- Could it contain text designed to manipulate AI into taking unintended actions?
- What sensitive data or action-capable tools are available to AI at the same time?
- Can external content be converted into constrained data before it influences tool use?
- Which permissions or tools should be unavailable while untrusted content is being processed?
- What actions should require confirmation even if external content appears to request them?
Confirming Whether External Actions Actually Occurred
- What action was supposed to happen?
- What evidence shows that the operation was actually submitted?
- Did the external system confirm completion or merely acknowledge receipt?
- Could the action have succeeded only partially?
- What is the current external state after the attempted action?
- Is there an identifier, record, receipt, log, or retrieved state that independently confirms success?
- Does the evidence show only that the action occurred, or also that it produced the intended outcome?
Evaluating Reliability and Uncertainty
Calibrating Confidence to Evidence
- How strong is the evidence behind this conclusion?
- Does the certainty of the wording match the strength of the evidence?
- Which claims are well established, and which remain tentative?
- What assumptions or missing information materially affect the conclusion?
- Is AI's expressed confidence supported by anything beyond fluent language?
- How consequential would it be if the conclusion were wrong?
- How should the answer change if its level of certainty were calibrated more closely to the evidence?
Tracing Claims to Their Basis
- Which claims matter enough to trace individually?
- What source, observation, calculation, tool result, model knowledge, or inference supports each one?
- Is the basis of the claim visible or merely implied?
- Which part comes directly from evidence, and which part comes from reasoning about that evidence?
- Are any claims resting only on assumptions or unsupported model knowledge?
- What additional support would be needed for the weakest important claims?
- How should unsupported claims be qualified, revised, or removed?
Checking Whether Citations Support Claims
- What exact claim is this citation supposed to support?
- What does the cited source actually establish?
- Does the source support the full claim or only part of it?
- Is the claim broader, stronger, more precise, or more general than the evidence allows?
- Does the source concern the same population, period, jurisdiction, definition, or context?
- Are several citations genuinely independent, or are they repeating the same underlying source?
- What should change if the citation does not support the statement as written?
Investigating Conflicts Between AI Answers and Trusted Sources
- What exactly conflicts between AI's answer and the source I trust?
- Is the disagreement factual, definitional, temporal, methodological, or interpretive?
- Could the source be outdated, incomplete, or valid only in a narrower context?
- What evidence is AI relying on for its conflicting answer?
- Could AI have misunderstood either my source or the retrieved evidence?
- What additional source or test could resolve the disagreement?
- What should remain uncertain if neither side can be established clearly?
Interpreting Evidence With Multiple Reasonable Conclusions
- What conclusions are genuinely compatible with the available evidence?
- Which interpretation has the strongest support, and why?
- What assumptions lead from the same evidence to different conclusions?
- Are the alternatives genuinely competing, or could several be true at once?
- Which possibilities are evidence-based and which are merely conceivable?
- What additional evidence would best distinguish among the leading interpretations?
- How should the result preserve legitimate uncertainty without treating every possibility as equally credible?
Challenging Unsupported Precision
- Which parts of the answer appear more precise than the evidence warrants?
- What evidence supports the exact number, date, probability, category, or ranking being stated?
- Is an estimate being presented as an exact value?
- Has AI converted a range or qualitative judgment into false numerical precision?
- Would a range, interval, category, or qualified statement better match the available evidence?
- Does the intended decision actually require this level of precision?
- What should be changed if the exact-looking result cannot be defended?
Distinguishing Source Content From AI Inference
- What does the source state directly?
- What has AI inferred from that source?
- What reasoning connects the evidence to the inferred conclusion?
- Which assumptions are required for that inference to hold?
- Could another reasonable interpretation fit the same source?
- Has AI presented an inference as though it were explicitly stated in the evidence?
- How should the final output make the boundary between source content and interpretation visible?
Verifying Claims in Unfamiliar Domains
- Which claims am I least able to evaluate independently?
- What kind of source would be authoritative for those claims?
- Can I verify them through primary evidence or a trusted specialist reference?
- Does understanding the evidence itself require domain expertise?
- What warning signs suggest AI may be generalizing beyond its competence?
- How consequential would an unnoticed specialist error be?
- Should a qualified expert review the conclusion before I rely on it?
Balancing Verification Effort Against Consequence
- What could happen if this claim or conclusion is wrong?
- How uncertain is it?
- How easy would an error be to detect without deliberate verification?
- How difficult would it be to reverse the consequences later?
- Which claims matter enough to verify individually?
- What is the cheapest reliable verification method available?
- At what point would additional checking cost more than the expected value of further certainty?
Deciding Under Incomplete Evidence
- What decision still needs to be made despite the missing information?
- What is known with enough confidence to use?
- What remains materially uncertain?
- Which assumptions are unavoidable if I act now?
- What additional information would actually have enough value to justify waiting for it?
- Could a reversible, staged, or adaptive decision reduce the cost of being wrong?
- What conclusion or commitment should remain limited because the evidence is incomplete?
Maintaining Human Judgment and Responsibility
Maintaining Independent Judgment
- Can I explain why this conclusion makes sense without simply repeating AI's reasoning?
- What did I think before seeing AI's answer?
- Which parts of my own reasoning have I effectively delegated?
- Do I understand the evidence and assumptions well enough to challenge the result?
- What credible reason could make AI's answer wrong?
- Would forming an independent view before asking for AI's recommendation reduce anchoring?
- What level of understanding do I need to retain before relying on the result?
Resolving Value and Tradeoff Decisions
- Which parts of this decision depend on facts, and which depend on values or priorities?
- What tradeoffs cannot be resolved by gathering more evidence?
- Whose interests, rights, or commitments are affected?
- What can AI clarify without deciding what should matter most?
- Which priorities belong to me, another accountable person, or a legitimate institution to determine?
- Is AI hiding a value judgment inside apparently neutral analysis?
- What choice must remain an explicitly human decision even after the analysis is complete?
Retaining Accountability for AI-Supported Work
- Who is ultimately responsible for this work or decision?
- Which parts were generated, analyzed, or influenced by AI?
- What level of review does that responsibility require?
- Can I explain and defend the final result if challenged?
- What errors would still be my responsibility even if AI introduced them?
- Does the work require disclosure, approval, attribution, or another formal accountability step?
- What must I personally verify before presenting or acting on the output under my authority?
- What sensitive information does this task involve?
- Does AI actually need all of it to complete the work?
- Could identifying details, unnecessary fields, or unrelated records be removed first?
- What organizational, contractual, professional, or legal rules govern this information?
- Where will the information be processed or stored?
- Could separate harmless details combine to reveal something sensitive?
- Can the task be completed with a less revealing representation of the information?
Retaining Human Authorization for External Actions
- What external action is AI proposing or preparing to perform?
- Who has legitimate authority to approve it?
- What consequences follow once the action occurs?
- Is the action difficult to reverse, externally visible, financially significant, or privacy-sensitive?
- What details should a human inspect before authorization?
- Which routine low-risk actions could reasonably operate under bounded prior approval?
- Where should case-by-case human authorization remain mandatory?
Preserving Learning When Using AI
- What capability am I trying to learn or strengthen through this task?
- Which cognitive work do I need to practice myself in order to develop that capability?
- Would having AI produce the answer immediately help learning or bypass it?
- What should I attempt before seeing AI's solution?
- Could AI act as a tutor, critic, hint provider, or practice partner instead of doing the whole task?
- How will I test whether I can perform the relevant reasoning independently afterward?
- What level of AI assistance improves progress without undermining the learning objective?
Avoiding Unnecessary Dependence on AI
- Which tasks am I becoming unable or unwilling to perform without AI?
- Do I still need those underlying capabilities for my role, resilience, or judgment?
- Can I detect and correct AI errors without asking AI again?
- What happens if AI becomes unavailable at an important moment?
- Is the dependence reducing a capability I still need to maintain?
- Which parts of the workflow are safe to depend on AI for because independent capability is no longer important?
- What level of dependence is acceptable given the consequences of failure?
Setting Boundaries When Expertise Is Insufficient
- What expertise is required to evaluate this result properly?
- Which parts can I judge reliably myself?
- What errors could sound plausible but remain invisible to me?
- Can authoritative sources compensate for the expertise I lack?
- Does interpreting those sources itself require specialist knowledge?
- Should AI's role be limited to explanation, preparation, or exploration rather than final judgment?
- When is qualified human review necessary before relying on the conclusion?
Exercising Human Judgment Over Effects on Others
- Who could be materially affected by this output or decision?
- What consequences might not be visible from my own perspective?
- Does AI have enough context about the people involved to assess those effects responsibly?
- What rights, interests, expectations, or vulnerabilities need explicit consideration?
- Which affected people possess information that should come from them rather than from AI inference?
- Does anyone need an opportunity to participate, respond, appeal, or correct the information being used?
- What responsibility remains human because other people will bear the consequences?
Resolving Conflicts Between AI Recommendations and Human Judgment
- What exactly does AI recommend, and where does my judgment differ?
- What evidence supports each position?
- What assumptions differ between the two conclusions?
- What relevant context or tacit experience do I have that AI may not possess?
- Could my own judgment be affected by habit, preference, or bias that AI is exposing?
- What external evidence, test, or expert view could help resolve the disagreement?
- Who should own the final decision if the disagreement remains genuinely unresolved?