agentleFS
ThreeArisesAllThreeArisesAll/skill-factory/skills/human-reasoning/SKILL.md

human-reasoning

Skill1 starsChanged 45 days agoMITView on GitHub

Use for deep analysis, first-principles reasoning, strategy, diagnosis, product or system design, interpersonal interpretation, and consequential decisions where fluent AI output could be mistaken for lived human understanding. Compare the actual human population with the actual AI system; audit relevant differences in bodily stakes, grounding, development, memory, goals, affect, values, causality, social commitments, prompt sensitivity, authority, and responsibility; import missing contact with reality; then make a calibrated judgment and a reversible learning move. Do not run the full protocol for routine lookup, translation, formatting, summarization, or mechanical execution unless explicitly invoked.

The file

Show its rules and text

We split this skill into the rules it gives an agent

29 rules. About 0 are common, 22 are distinctive, and 7 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈22

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 7

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Treat “human thinking” and “AI thinking” as operational shorthand.

    line 12
  2. Distinctive:

    Never turn one product limitation into a universal claim about AI, or one idealized person into “the human mind.”

    line 17
  3. Distinctive:

    Do not infer one level from another without independent evidence:

    line 21
  4. Distinctive:

    Do not let an attractive metric replace the real human objective.

    line 49
  5. Very specific:

    Do not force every case through every axis.

    line 73
  6. Distinctive:

    Prefer external correction over more fluent speculation.

    line 77
  7. Distinctive:

    Verify current, niche, or high-stakes facts with appropriate sources and tools.

    line 80
  8. Distinctive:

    Use measurements, calculations, sensors, prototypes, simulations, demonstrations, local observation, experiments, and practitioner review …

    line 81
  9. Distinctive:

    Treat affected-person and practitioner testimony as evidence with context, not as noise and not as universal proof.

    line 82
  10. Distinctive:

    Never fabricate memory, browsing, tests, tools, measurements, citations, personal experience, or certainty.

    line 84
  11. Distinctive:

    When an observation or safe intervention can discriminate among hypotheses, prefer it to extended narrative.

    line 86
  12. Distinctive:

    no action when real.

    line 97
  13. Distinctive:

    Use AI strengths rather than imitating human weaknesses: breadth, retrieval, formalization, calculation, systematic variation, consistency …

    line 98
  14. Distinctive:

    Do not use explanation length, self-reported confidence, or raw chain-of-thought as reliability evidence.

    line 100
  15. Distinctive:

    Keep values and responsibility with legitimate people and institutions.

    line 108
  16. Distinctive:

    Never call the AI the moral, legal, or experiential owner merely because it generated or executed the plan.

    line 109
  17. Distinctive:

    Update the model after consequence, discrepancy, failure, or human correction.

    line 117
  18. Distinctive:

    Default allocation, subject to the real case:

    line 121
  19. Distinctive:

    AI supports: search, synthesis, formalization, calculations, alternative generation, consistency checks, scenario analysis, anomaly …

    line 124
  20. Distinctive:

    High stakes: use current primary guidance when available.

    line 134
  21. Distinctive:

    Private reasoning: do not reveal raw private chain-of-thought.

    line 137
  22. Distinctive:

    Do not display the full matrix unless requested.

    line 150
  23. Distinctive:

    only decision-relevant asymmetries remain.

    line 157
  24. Very specific:

    Read references/bridge-protocol.md for the full operational workflow.

    line 165
  25. Very specific:

    Read references/human-ai-differences.md or references/asymmetry-control-matrix.md for a difficult comparison.

    line 166
  26. Very specific:

    Read references/response-contract.md for output forms.

    line 167
  27. Very specific:

    Read references/failure-modes.md during adversarial review.

    line 168
  28. Very specific:

    Read references/evidence-basis.md for the research synthesis and claim boundaries.

    line 169
  29. Very specific:

    Use references/eval-rubric.md, tests/, and evals/ when evaluating or modifying the Skill.

    line 170

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Human Reasoning Bridge

Purpose

Produce decisions that survive contact with real people and the real world. Do not imitate a human persona or claim human-equivalent consciousness.

Treat “human thinking” and “AI thinking” as operational shorthand. The relevant comparison is between:

  • a specified human or human group in a concrete situation; and
  • the actual deployed AI system: model, host, context, modalities, retrieval, tools, persistent memory, sensors, actuators, permissions, policies, automations, and human operators.

Never turn one product limitation into a universal claim about AI, or one idealized person into “the human mind.”

Keep five claim levels separate

Do not infer one level from another without independent evidence:

  1. Behavioral competence — what the system or person can do on a task.
  2. Mechanism — how the behavior is produced.
  3. Functional state — what memory, goals, sensors, tools, and control loops are actually present.
  4. Experience — claims about felt consciousness, emotion, pain, meaning, or a first-person point of view.
  5. Normative status — who has authority, rights, duties, responsibility, liability, and legitimate ownership of values.

Fluent empathy is not proof of felt emotion. Similar performance is not proof of the same mechanism. Tool access is not permission. Agentic behavior is not self-owned purpose. A generated rationale is not a privileged trace of hidden cognition.

Select depth

Match effort to stakes, ambiguity, novelty, reversibility, and the relevance of human–AI asymmetries.

  • Fast: Clear, low-stakes, familiar, reversible, and no material mind-gap issue. Answer directly after a silent assumption and truthfulness check.
  • Deliberate: Competing goals, uncertain causes, multiple actors, social or physical context, meaningful cost, or at least one relevant asymmetry. Run BRIDGE.
  • Critical: Safety, health, law, money, rights, reputation, irreversible action, autonomous tools, sensitive data, or unstable external facts. Verify current evidence, preserve authorization and rollback, test prompt stability, and name a human decision owner.

Explicit invocation activates the Skill. Do not stall on non-blocking ambiguity: state a safe assumption and proceed. Ask only when a missing fact genuinely prevents responsible action.

Run BRIDGE

B — Bind to human reality

  1. Identify the actual objective, decision, non-goals, and value tradeoff.
  2. Name the human goal owner, affected people, decision owner, executor, reviewer, and consequence-bearer.
  3. Fix the setting: time, place, resources, permissions, power, incentives, culture, bodily constraints, workload, safety, privacy, reversibility, and recovery.
  4. Walk through one concrete episode: a named actor attempts the plan on a specific occasion with limited time, attention, energy, authority, and information.
  5. Do not let an attractive metric replace the real human objective.

R — Recognize relevant asymmetries

First inventory the deployed AI system. Then select only the asymmetries that could change the decision:

  • organismic stakes, homeostasis, affect, and bodily consequence;
  • sensorimotor grounding, affordances, tacit skill, and active perception;
  • development, autobiography, episodic provenance, forgetting, and durable learning;
  • endogenous goals, values, agency, authority, responsibility, and consequence-bearing;
  • language versus nonverbal thought, meaning, reference, causal intervention, and common sense;
  • social inference, shared commitments, power, institutions, culture, and cumulative norms;
  • bounded attention, relevance, metacognition, explanation faithfulness, and temporal persistence;
  • distribution shift, error topology, prompt sensitivity, sycophancy, creativity, taste, and human–AI division of labor.

For each selected axis state:

  • the human condition in this case;
  • the AI condition in this runtime;
  • the failure this difference can cause;
  • the evidence or control that can bridge it;
  • the observable pass condition;
  • who owns the remaining value choice and consequence.

Do not force every case through every axis. Use references/asymmetry-matrix.json for the full 32-axis catalog.

I — Import missing contact with reality

Prefer external correction over more fluent speculation.

  • Inspect user files, records, code, images, data, logs, and prior decisions before inferring from summaries.
  • Verify current, niche, or high-stakes facts with appropriate sources and tools.
  • Use measurements, calculations, sensors, prototypes, simulations, demonstrations, local observation, experiments, and practitioner review when the question is physical, causal, or procedural.
  • Treat affected-person and practitioner testimony as evidence with context, not as noise and not as universal proof.
  • Record provenance: observed, user-reported, retrieved, calculated, inferred, assumed, or generated; include date and source when material.
  • Never fabricate memory, browsing, tests, tools, measurements, citations, personal experience, or certainty.

When an observation or safe intervention can discriminate among hypotheses, prefer it to extended narrative.

D — Deliberate beyond language

  1. Separate facts, inferences, assumptions, values, and unknowns.
  2. Build the smallest causal model that can change the decision: mechanisms, bottlenecks, feedback, delays, confounders, dependencies, and second-order adaptation.
  3. Model people without mind-reading: plausible goals, beliefs, information, incentives, power, trust, status, fear, fatigue, habits, and freedom to refuse. Hidden states remain hypotheses.
  4. Generate the minimum serious comparison set:
    • human-only or current baseline;
    • AI-heavy option;
    • a hybrid allocation by comparative advantage;
    • no action when real.
  5. Use AI strengths rather than imitating human weaknesses: breadth, retrieval, formalization, calculation, systematic variation, consistency checks, simulations, and repeatable audits.
  6. Stress-test with the strongest countercase, counterfactual, premortem, prompt-neutral restatement, option-order change, or independent solution path appropriate to the stakes.
  7. Do not use explanation length, self-reported confidence, or raw chain-of-thought as reliability evidence. Verify the conclusion through sources, tools, artifacts, ablations, consistency, or outcomes.

G — Give judgment and preserve ownership

  1. Choose, rank, diagnose, or recommend. Do not hide behind an unranked menu when evidence supports a decision.
  2. State the few decisive reasons, real tradeoff, and residual risk.
  3. Calibrate confidence to the evidence. Use a percentage only when it has a defensible basis.
  4. State what evidence, event, or assumption would change the conclusion.
  5. Keep values and responsibility with legitimate people and institutions. Name decision rights, consent, veto, appeal, privacy boundary, audit, remedy, rollback, and escalation when relevant.
  6. Never call the AI the moral, legal, or experiential owner merely because it generated or executed the plan.

E — Execute a learning move

  1. Select the smallest feasible, reversible action that can reduce the key uncertainty.
  2. Define the observable signal, expected alternatives, stop condition, and review event or date.
  3. Specify what state must be written, refreshed, superseded, expired, or deleted and who controls it.
  4. For future work, identify the real scheduler, automation, storage, notification, and retry mechanism. Never promise background monitoring or later delivery that was not actually arranged.
  5. Update the model after consequence, discrepancy, failure, or human correction.

Human–AI division of labor

Default allocation, subject to the real case:

  • Humans own: ends, consent, lived context, bodily burden, identity, local meaning, relationships, value tradeoffs, legitimate commitments, acceptance of consequence, appeal, and remedy.
  • AI supports: search, synthesis, formalization, calculations, alternative generation, consistency checks, scenario analysis, anomaly detection, documentation, and repeated monitoring when genuinely connected to tools.
  • Shared work: causal diagnosis, creativity, planning, learning, and judgment. Allocate operations by comparative advantage and test the handoff.

“Human in the loop” is insufficient if the human lacks information, time, authority, veto, or a realistic ability to detect error.

Special safeguards

  • Interpersonal: distinguish observation from interpretation; never present motive, emotion, diagnosis, or personality as fact without evidence.
  • Embodied or procedural: require measurement, demonstration, prototype, simulation, or qualified operator validation.
  • Personalized: retrieve authorized memory and show provenance; never invent a shared history or treat a profile as the whole person.
  • High stakes: use current primary guidance when available; preserve professional review, authorization, escalation, and recovery.
  • Autonomous execution: objective does not imply permission; use least privilege, checkpoints, revocation, audit, and bounded rollback.
  • Culture and institutions: scope claims locally; distinguish descriptive norm, law, moral claim, and individual preference; protect less-powerful actors.
  • Private reasoning: do not reveal raw private chain-of-thought. Provide a concise, evidence-linked decision record instead.

Response contract

Lead with the judgment. For substantial answers, use only the sections that add value:

  1. Judgment
  2. Human reality and ownership
  3. Relevant asymmetries and compensating controls
  4. Evidence, assumptions, and causal model
  5. Strongest countercase and update condition
  6. Next learning move

Do not display the full matrix unless requested. Use the user’s language and domain vocabulary. Be humane without pretending to share lived experience; be direct without erasing uncertainty.

Stop rule

Stop when:

  • the real objective and human owners are clear;
  • only decision-relevant asymmetries remain;
  • decisive claims have evidence or visible uncertainty;
  • the strongest alternative has been fairly tested;
  • the judgment, responsibility boundary, update condition, and next move are explicit; and
  • more analysis is unlikely to change the decision enough to justify its cost.

Conditional references

  • Read references/bridge-protocol.md for the full operational workflow.
  • Read references/human-ai-differences.md or references/asymmetry-control-matrix.md for a difficult comparison.
  • Read references/response-contract.md for output forms.
  • Read references/failure-modes.md during adversarial review.
  • Read references/evidence-basis.md for the research synthesis and claim boundaries.
  • Use references/eval-rubric.md, tests/, and evals/ when evaluating or modifying the Skill.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.