--- name: graphical-user-interface description: "Design/review HCI across GUI/TUI/CLI/web/desktop/mobile. Use for flows, models, feedback, responsiveness, a11y, errors, medium fit. Not styling/components." --- # Human-computer interaction Design the relationship between human intent and machine behavior, not a collection of UI elements. Apply principles proportionally; surface consequential trade-offs, not a mandatory design ceremony. ## Establish context - Identify intended users, tasks, frequency, stakes, environment, and project goals before choosing priorities. - Inspect relevant project contracts, existing interfaces, and observed workflows before asking questions they can answer. - Treat precedent as evidence, not authority; challenge inherited interaction when task benefits justify relearning and migration costs. - Ask when consequential priorities remain unresolved; disclosed assumptions are not evidence. - Let context decide novice/expert emphasis and acceptable learning cost; do not impose one universal audience or simplicity ideal. - Treat conflicts between project goals and informed choice as explicit project decisions; this skill is not standing authorization for misleading behavior. ## Fit the medium - Preserve task intent across media; adapt interaction rather than transplanting another medium's model or brute-forcing its limitations. - Assess actual input, output, persistence, accessibility, connectivity, surrounding tools, and resource constraints, not just platform labels or viewport size. - Exploit strengths that help the task; do not require features merely because the medium supports them. - Respect useful host capabilities and conventions; consistency means transferable expectations, not identical presentation everywhere. - Support equivalent core-task completion through accessible alternatives; equivalent need not mean identical interaction. ## Minimize whole-task effort - Optimize completion, understanding, search, memory, navigation, errors, and repeated-use effort together; neither density nor minimalism wins by default. - Make available actions, scope, current state, and likely consequences discoverable where needed. - Match language and information organization to users' tasks and mental models, not implementation structure. - Remove incidental complexity; decide contextually when to teach necessary concepts versus abstract them away. - Support progression toward mastery without making recall or hidden shortcuts the only usable path. - Disclose secondary complexity when helpful without hiding frequent operations, necessary comparisons, or material consequences. - Preserve orientation and useful context across interruption; require attention only when it helps the task. ## Respect both sides of the interface - Give human usability and mechanical sympathy equal design standing: account for CPU, memory, storage, bandwidth, energy, scheduling, and communication costs. - Keep semantic state structured and machine-suitable; derive task-specific human views rather than parsing presentation back into application truth. - Preserve precision, identity, units, and meaning internally even when a human view rounds, summarizes, localizes, or reorders them. - Choose internal and exchange representations for their actual workloads; machine-readable does not mean one prescribed serialization everywhere. - Expose stable machine interfaces when workflows, integration needs, or project contracts warrant them; do not add external automation solely because internal state is structured. - Treat latency, throughput, working-set size, and energy as interacting costs; improving one does not prove overall improvement. ## Make time part of the design - Prefer responsiveness; establish resource exceptions from domain, app, project, and relevant existing interfaces, then ask if unresolved. - Separate input acknowledgment, useful result, and durable completion; immediate acknowledgment must not imply work is finished. - Remove avoidable waiting before decorating it; keep unnecessary computation, transfers, redraws, and blocking work outside the interaction's critical path. - Consider incremental work, locality, caching, or speculation only with their correctness and resource costs; preloading can reduce waiting, not guarantee instant results. - Keep useful work available during unavoidable delays; provide truthful state and meaningful cancellation, safe exit, or resumption according to stakes and workflow. - Distinguish pending, locally retained, synchronized, failed, and durably completed states wherever those differences affect user decisions. - Do not invent progress, promise unsupported completion times, or hide stale results as current truth. - Set timing expectations for the actual interaction; continuous manipulation, discrete actions, and long jobs need different treatment. ## Support error and recovery - Prevent costly slips and misunderstandings without making every action a confirmation ritual. - Make scope and irreversible consequences understandable before commitment; support correction and reversal where appropriate. - Preserve useful work through errors and interruption; establish unknown retention stakes before weakening protection against potentially irreversible loss. - A justified exception for explicitly disposable data does not settle unknown stakes, accessibility, or other agency trade-offs. - Explain failure in task terms, preserve valid input, and offer an actionable recovery path. ## Ground decisions - Connect recommendations to task evidence, project constraints, or established principles; identify which support each claim. - Use reversible prototypes when representative observations are unavailable; predicted benefits remain hypotheses until tested. - Evaluate representative tasks and relevant failure states, including constrained environments; inspect completion, effort, mistakes, recovery, accessibility, and perceived control. - Measure input-to-feedback, useful-result latency, completion, and resource costs relevant to the decision; averages or throughput alone can hide interaction stalls. - Report observed results and remaining uncertainty; profile, verify rendered fidelity, and test as the change warrants. ## Research Read [research and interpretation](references/research.md) when source authority, evidence limits, medium-specific application, or a contested trade-off needs explanation. Keep frameworks, component catalogs, visual style rules, and platform implementation recipes outside this skill.