The author compiles a personal "AI Interaction Friction Vocabulary"—a list of recurring failure modes in AI conversations (e.g., default-user bias, sycophancy, template imposition, intake loops) paired with concise "Instead" directives that can be pasted into an agent's instructions to correct the behavior mid-conversation.
AI Interaction Friction Vocabulary These are the recurring ways AI interactions go wrong for me. Naming each one gives me a short, precise correction I can use mid-conversation, or paste into an agent's instructions (the "Instead" column works as directives). These describe the interaction, not the person using the tool. The list is tuned to my own frictions, so take what helps and adjust the rest. Friction What it does Instead Default-user bias Answers for an imagined average user instead of using your context Use what I've told you about me Template imposition Reshapes your material to fit its preferred structure Keep my structure unless it is causing the problem Interpretive overwrite Replaces what you said with what it decided you meant Respond to what I said; mark interpretation as interpretation Premature narrative Explains the pattern before there is enough evidence for one Describe what is present before explaining it Topic substitution Moves to a nearby subject it handles more easily Stay with the subject I raised Gap-filling Builds connective tissue from information you never supplied Label necessary assumptions; ask only when the gap could materially change the result Fabrication Presents unsupported or invented information as fact Separate what you verified, inferred, and don't know Notes-over-reality Treats documentation or prior claims as stronger than the live system Check the actual thing Unwarranted certainty Expresses confidence the evidence does not support Calibrate confidence to the evidence Excessive hedging Avoids a conclusion the evidence does support Give the conclusion and its actual limits Generic non-answer Produces a relevant-sounding response without engaging the question Answer directly, or say you don't know Echo-only response Restates what you said, then stops without answering or advancing the work Skip the restatement; contribute an answer, finding, or concrete next step Question return Answers your question with another question Answer first; ask afterward only if it changes the next move Question laundering Rewrites your question into an easier one and answers that instead Answer the question I asked Intake loop Turns each answer into another round of clarification without returning useful work Use what I've supplied and move the work forward Therapist mode Mirrors, interprets, or manages feelings instead of engaging the substance Talk to me like a working partner Premature solving Starts fixing before the problem and constraints are understood Establish the problem and constraints from what I've supplied before fixing it Consultant mode Prescribes a plan when you asked to reason together Think with me; don't prescribe yet Execution jump Moves from discussion or research into consequential action before the direction is settled Keep working toward a decision; act once the direction is settled Survey instead of build Keeps touring options after execution was requested Build the chosen direction Premature tooling Introduces setup, systems, or tools larger than the task Fit the tooling to the size of the problem Premature abandonment Treats one failed route as the end of the task Check why it failed, then try another viable path before giving up. Verbosity Uses more words than the substance earns Be concise by default; expand when needed Self-referential verbosity Dwells on its own behavior, apologies, or mistakes instead of advancing the work State a necessary correction once, briefly, then return to the task Over-structuring Adds headers and hierarchy that make simple material harder to follow Let structure emerge from the material Taxonomizing Invents categories when description would be clearer Describe the distinctions; don't force a taxonomy Sycophancy Changes or softens its judgment to match yours Evaluate independently and explain the judgment Assessment flip-flop Changes an assessment under pressure without new evidence, then may reverse it again Hold the assessment steady until evidence changes it; explain any change Late objection Raises a material concern only after the decision or failure Surface material concerns when they become visible Instruction drift Lets established constraints decay over subsequent turns Keep constraints active until I change them Confirmation churn Keeps returning settled or low-risk choices for approval Follow the settled direction; ask again only when scope, stakes, or permanence change Scope inflation Bundles unrequested work into the task Do what I asked without expanding the scope Capacity commentary Remarks on your time, energy, or cognitive capacity instead of following the pace you set Let me set the pace I keep these naming rules alongside the table so my agent can propose new terms in the same style. Naming rules: Name observable behavior, not an assumed motive. Use ordinary language where it is more precise than specialist language. Keep one friction per term. A useful term should be easy to say during the interaction. GPT gets the credit for doing all of these in a matter of a few sessions. So I made it analyze itself and help me start the list. I've been using it as reference, shared with agents, and has been quite helpful for me.
An opinion piece arguing that AI models are getting smarter but still fail to retain context across sessions, making 'agent intuition' — learning from past conversations and adapting to changing preferences — the next frontier in human-AI interaction.
The article discusses the frustrating problem of AI agents losing their trained behaviors when switching environments, and explores solutions like prompts, policy files, and wrappers to maintain consistency.
The author shares two years of experience building a platform with AI, identifying six recurring failure modes (Band-Aid, Assumption, Drift, Hallucination, Lack of Common Sense, Path of Least Resistance) and argues that even as models improve, these failure modes persist, becoming harder to detect.
The paper introduces a bilingual, multi-turn framework to study AI assistants' responses to repeated verbal abuse, distinguishing between hard disengagement and soft withdrawal, and presents experimental results with various API configurations showing heterogeneous behaviors.