1---2name: ai-agent-ux3description: Use when an AI agent or copilot interface needs status, plans, approvals, uncertainty, interruption, or trust controls across web or mobile. Do not use for formatting a single AI answer; route output structure and citations to ai-output-design.4---56# AI Agent UX7Acknowledgement: Shared by Peter Bamuhigire, techguypeter.com, +256 784 464178.89<!-- dual-compat-start -->1011## Use When1213- Design agentic UI patterns, approval moments, progress status, uncertainty handling, and explainability surfaces.14- Adapt agent UX for mobile and web workflows without exposing raw JSON or implementation detail.15- Shape user trust, control, recovery, and human handoff in AI agent experiences.1617## Do Not Use When1819- The work is not AI-specific or agentic-AI-specific.20- A narrower retained AI parent skill fits the request better.2122## Required Inputs2324| Input | Source | Evidence |25|---|---|---|26| User goal, agent powers, and prohibited actions | Product owner and capability policy | Enumerated actions and permission boundary |27| Agent lifecycle, failure modes, and latency | Engineering contract | State/event list and recovery paths |28| Human approval and audit requirements | Risk, legal, or operations owner | Approval thresholds and log requirements |2930- Product, tenant, user, data, risk, and operational context relevant to the AI workflow.31- Target artifact: design, implementation plan, audit, test strategy, UX flow, commercial policy, or runbook.32- Constraints from security, privacy, reliability, billing, support, and compliance stakeholders when relevant.3334## Workflow35361. Read this SKILL.md first.372. Load `references/routing.md` to select the absorbed child reference that matches the task.383. Load only the selected child reference files needed for the current request.394. Produce execution-oriented output with assumptions, risks, evidence, and next actions where relevant.4041## Decision Rules4243| Condition | Choice | Wrong-choice failure |44|---|---|---|45| Action is consequential or irreversible | Preview plus explicit approval | Silent autonomy causes harm and destroys trust |46| Work is long-running | Persistent progress with pause/cancel | Spinner-only feedback looks stalled and removes control |47| Confidence is uncertain | Show evidence, limits, and next verification | Invented certainty misleads users |4849## Capability Contract5051- Must inspect capability and event contracts; review is read-only unless implementation is requested.52- May prototype or edit in-scope UI, but may not grant the agent new powers, execute consequential actions, or publish without explicit authority.5354## Degraded Mode5556- If powers or approval thresholds are unknown, stop the interaction specification and return the missing-policy questions.57- Without a runnable agent, produce a state/event prototype marked simulated. Recover failed states with retry, edit, cancel, or human escalation and retest the path.5859## Quality Standards6061- Keep routing explicit: name which reference files were used when the work depends on absorbed material.62- Preserve tenant isolation, auditability, cost controls, safety gates, and operational evidence when they matter.63- Prefer concrete contracts, checklists, tables, schemas, runbooks, and decision records over broad summaries.6465## Anti-Patterns6667- Loading every absorbed reference by default.68- Treating AI-specific billing, compliance, safety, or UX concerns as generic SaaS work without checking AI failure modes.69- Hiding retired skill names; old slugs must remain discoverable through `references/routing.md`.70- Hiding plans or actions behind a personality animation. Correction: expose task, state, evidence, and controls.71- Auto-executing a consequential step after ambiguous consent. Correction: preview scope and require explicit approval.72- Showing an endless spinner with no cancel path. Correction: persist progress, elapsed state, pause, and cancellation.7374## Outputs7576| Output | Consumer | Evidence and acceptance |77|---|---|---|78| Agent interaction specification | Product and engineering | State, approval, interruption, error, and recovery paths are explicit |79| Trust/control review | Risk and QA | Consequential actions have preview, consent, audit, and cancellation evidence |8081- A concrete deliverable matched to the request: architecture, implementation plan, audit, policy, runbook, UX flow, test strategy, or operating model.82- The selected consolidated reference files and any assumptions, risks, evidence requirements, or follow-up actions that affect execution.83## References8485- `doctrine/design-doctrine.md` — the anti-slop charter; agent UX must avoid "AI hammer looking for a nail" feature-cramming.86- `doctrine/references/ai-slop-taxonomy.md` — the product/interface slop tells (AI feature where search was faster; ungrounded output; "AI" badges as decoration) that agent surfaces must not exhibit.87- `ai-output-design` — sibling skill for the AI output surface (verifiability, grounding, forward-actions) these agent flows render into.88- `references/routing.md` maps retired child skill slugs to their consolidated reference folders.89- `references/generative-ai-ui-ux.md` covers generative-AI prompt, streaming, confidence, source, review, recovery, and human-override UI patterns.9091## Book-derived additions9293For agency, cognitive-privacy, memory, refusal, correction, and exit controls,94load [human agency and cognitive privacy](references/human-agency-and-cognitive-privacy.md).9596## Consolidated Child References9798- Load `references/routing.md` to map retired AI child skill slugs to their reference modules.99100## Examples101102- `examples/agent-ux-spec-worked.md` — a full worked agent-UX spec for a sample copilot ("Ledger"): status surface, approval/confirmation UX, uncertainty + confidence display, interruption/stop, error recovery, and the human-in-the-loop gates that tie them together. Use as a fill-in template.103<!-- dual-compat-end -->
peterbamuhigire