# RAD Design System > A design system for AI surfaces that act, decide, and operate autonomously. Built by Jackie Curry. RAD (Responsible · Accountable · Disclosed) is a design system for agentic AI user interfaces. It provides standardized components, behavioral contracts, and interaction patterns for the specific trust failures that occur when AI systems communicate with and act on behalf of humans. RAD is used when building: agent execution UIs, AI copilot interfaces, LLM output surfaces, human-in-the-loop workflows, AI-powered dashboards, and audit and governance interfaces. **Author:** Jackie Curry **Site:** https://jackiecurry.github.io **Version:** 1.0 · © 2025 Jackie Curry. All Rights Reserved. --- ## Core Principles **R — Responsible.** Every AI action is attributable. Systems surface what they did, on what basis, and under whose authority. **A — Accountable.** Human oversight is preserved. Consequential decisions require explicit human approval before execution. **D — Disclosed.** AI involvement is never hidden. Systems identify themselves, their data sources, and their confidence levels. --- ## RAD Design System Overview - [Overview](https://jackiecurry.github.io/rad.html): Introduction to the RAD design system, its principles, and component groups. - [AI Components](https://jackiecurry.github.io/rad-components.html): Full catalog of all 34 components across 7 groups. - [AI Patterns](https://jackiecurry.github.io/rad-patterns.html): Multi-component composition patterns for complete AI interaction flows. - [Design Tokens](https://jackiecurry.github.io/rad-tokens.html): Color, typography, and spacing token definitions. - [Theming](https://jackiecurry.github.io/rad-theming.html): How to apply and customize RAD themes. - [Downloads](https://jackiecurry.github.io/downloads.html): RAD kits for Claude Code and Figma. --- ## Component Groups ### Transparency & Disclosure Components that make AI behavior legible — what the system did, why, and on what basis. - [RAD-01 · Disclosure Alerts](https://jackiecurry.github.io/rad/components/disclosure-alerts.html): Alerts users when an AI is acting or has acted on their behalf. Prevents hidden autonomous action. - [RAD-03 · Transparency Popovers](https://jackiecurry.github.io/rad/components/transparency-popovers.html): Explains why the AI made a specific decision, on demand. Surfaces data sources, model logic, and confidence. - [RAD-04 · Bias Check Prompts](https://jackiecurry.github.io/rad/components/bias-check-prompts.html): Warns when AI output carries statistical risk, data limitation, or potential bias. - [RAD-07 · Uncertainty Indicators](https://jackiecurry.github.io/rad/components/uncertainty-indicators.html): Communicates model confidence level at the point of output. Distinguishes high-confidence from uncertain results. - [RAD-18 · Confidence Threshold Warning](https://jackiecurry.github.io/rad/components/confidence-threshold.html): Surfaces when output falls below an operator-defined confidence level. - [RAD-19 · Algorithmic Nudge Disclosure](https://jackiecurry.github.io/rad/components/algorithmic-nudge-disclosure.html): Discloses when AI is ranking, pre-selecting, or framing choices on the user's behalf. - [RAD-11 · Streaming Text Display](https://jackiecurry.github.io/rad/components/streaming-text.html): Patterns for token-by-token LLM output rendering. - [RAD-12 · AI Loading States](https://jackiecurry.github.io/rad/components/loading-states.html): Visual states while the model is reasoning, retrieving, or generating. - [RAD-14 · AI Error States](https://jackiecurry.github.io/rad/components/error-states.html): Handles timeout, refusal, hallucination detection, and rate limit errors. - [RAD-15 · Session Management](https://jackiecurry.github.io/rad/components/session-management.html): Surfaces context window usage, memory state, and session history. ### Human Control & Oversight Components that preserve human agency over AI decisions and actions. - [RAD-02 · Human-in-the-Loop Controls](https://jackiecurry.github.io/rad/components/hitl-controls.html): Intercepts agent execution before irreversible or high-consequence actions. Requires explicit human approval before proceeding. - [RAD-06 · Impact Assessment Toggles](https://jackiecurry.github.io/rad/components/impact-assessment.html): Surfaces the footprint of agent runs — what systems will be touched, what data will be accessed. - [RAD-08 · Agent State Indicators](https://jackiecurry.github.io/rad/components/agent-state.html): Shows when an agent is actively running a multi-step task and what phase it is in. - [RAD-09 · Consent & Scope Gates](https://jackiecurry.github.io/rad/components/consent-scope-gates.html): Requires user consent when an agent requests access to a new system or data source. - [RAD-21 · Agent Attention Triggers](https://jackiecurry.github.io/rad/components/agent-attention-triggers.html): Alerts when a monitoring system detects an anomaly, scope breach, or decision fork requiring human attention. - [RAD-10 · Recovery & Override Controls](https://jackiecurry.github.io/rad/components/recovery-override.html): Surfaces when an agent has errored, paused, or produced unexpected output. Provides override, rollback, and escalation controls. - [RAD-16 · Feedback & Correction](https://jackiecurry.github.io/rad/components/feedback-correction.html): Lets users signal output quality and correct AI responses inline. ### Accountability & Audit Components for logging, tracking, and surfacing the environmental and organizational cost of AI. - [RAD-05 · Audit Trail Widgets](https://jackiecurry.github.io/rad/components/audit-trail.html): Logs agent actions with timestamps, attribution, and reversibility status for accountability. - [RAD-20 · Environmental Impact Indicators](https://jackiecurry.github.io/rad/components/environmental-impact.html): Surfaces the energy, water, and carbon cost of AI inference at point of use. ### Foundation Core input and empty-state components for AI interfaces. - [RAD-13 · Prompt Input](https://jackiecurry.github.io/rad/components/prompt-input.html): Primary text input component for AI interaction. Handles single-line and multi-line prompt entry. - [RAD-17 · Empty & First-Run States](https://jackiecurry.github.io/rad/components/empty-states.html): Zero state and first-time user onboarding for AI interfaces. - [AI Artifacts](https://jackiecurry.github.io/rad/components/ai-artifacts.html): Display component for AI-generated documents, code, charts, and images. ### Capability Discovery Components that help users understand what an AI interface can do. - [Suggested Prompts](https://jackiecurry.github.io/rad/components/suggested-prompts.html): Surfaces example actions and questions to help users understand what to ask. Addresses the empty prompt box problem. - [Suggested Next Actions](https://jackiecurry.github.io/rad/components/suggested-next-actions.html): Post-response recommendations that guide users toward logical next steps in a workflow. - [Prompt Enhancement](https://jackiecurry.github.io/rad/components/prompt-enhancement.html): Shows a rewritten version of the user's prompt that would produce better results. Enhancement is always opt-in — never automatic. - [Task Builder](https://jackiecurry.github.io/rad/components/task-builder.html): Structured form controls that help users compose multi-step agent instructions without writing complex prompts. ### Agent Activity Components for surfacing what an agent is doing and has done. - [Agent Activity Timeline](https://jackiecurry.github.io/rad/components/agent-activity-timeline.html): Chronological view of agent steps with status indicators — complete, error, active, and pending states. - [Tool Execution Log](https://jackiecurry.github.io/rad/components/tool-execution-log.html): Real-time log of tool calls made by an agent, including type, target, and result. - [Collapsible Agent Steps](https://jackiecurry.github.io/rad/components/collapsible-agent-steps.html): Condenses long agent reasoning chains into expandable summaries. Preserves transparency without overwhelming the user. - [Process vs Result Layout](https://jackiecurry.github.io/rad/components/process-result-layout.html): Separates agent reasoning from final output. Result is always primary; process is always auditable. ### Context Management Components for managing what information an AI has access to in a session. - [Context Sources](https://jackiecurry.github.io/rad/components/context-sources.html): Shows what data sources, files, and connections the AI is drawing from in the current session. - [Context Pills](https://jackiecurry.github.io/rad/components/context-pills.html): Compact inline indicators showing active context items attached to the current prompt. - [Active Memory Panel](https://jackiecurry.github.io/rad/components/active-memory-panel.html): Surfaces what the AI has stored in memory across sessions and lets users manage those entries. - [Context Scope Selector](https://jackiecurry.github.io/rad/components/context-scope-selector.html): Lets users define the scope of an AI query — current file, folder, project, or custom. --- ## Key Concepts **Agentic AI interface:** A user interface where an AI system takes autonomous actions — writing data, sending communications, executing transactions — rather than only providing information. **Human-in-the-loop (HITL):** A design pattern requiring explicit human approval before an AI executes a consequential action. Never auto-approve; auto-cancellation only. **Disclosure:** The principle that AI involvement is never hidden. Users must always know when AI is acting, what data it used, and what confidence level applies. **Behavioral contract:** In RAD, each component defines not just visual structure but behavioral rules that apply regardless of theming — what the component must do, must never do, and must communicate. **Transparency popover:** A UI element that surfaces why an AI made a specific decision on demand, without cluttering the primary interface. **Confidence threshold:** An operator-defined minimum confidence level below which AI output must be flagged to the user before being acted upon. **Algorithmic nudge:** When an AI system ranks, pre-selects, or frames choices in a way that influences the user's decision. RAD requires this to be disclosed. **Context window:** The total amount of text (tokens) an LLM can process in a single session. RAD's Session Management component surfaces context window usage to users. --- ## Design Philosophy RAD was built because AI systems are increasingly taking actions — not just answering questions. The failure modes of agentic AI are different from the failure modes of traditional software: - Users don't know an agent acted until damage is done - Users can't tell the difference between confident and uncertain AI output - Users have no mechanism to audit what an agent did or why - Users can't stop an agent that is heading in the wrong direction RAD addresses these failure modes through standardized components that make AI behavior legible, controllable, and auditable — at the point of interaction, not in a log file reviewed later. The three principles — Responsible, Accountable, Disclosed — are not aspirational. They are implementation contracts.