| name | ai-output-design |
| description | Use when an AI answer surface needs structure, citations, grounding, editable refinement, or a chat-versus-canvas decision. Do not use for agent autonomy, approvals, or execution status; route those interaction controls to ai-agent-ux. |
| metadata | {"portable":true,"category":"04-web-and-ui-design","compatible_with":["claude-code","codex"]} |
AI Output Design
Acknowledgement: Shared by Peter Bamuhigire, techguypeter.com, +256 784 464178.
Use When
- Designing any AI output surface — chat bubbles, result panels, document drafts, canvas artifacts.
- The deliverable needs structure, verifiability, grounding, forward-actions, or inline refinement.
Do Not Use When
- Designing the chrome around AI (loading, streaming, confidence states on input side) — use
ai-agent-ux (its ai-ux-patterns reference module).
- Designing long-running agent control primitives — use
ai-agent-ux (its ai-agentic-ui reference module).
Required Inputs
| Input | Source | Evidence |
|---|
| User task and output decisions | Product brief or observed workflow | Ranked decisions and follow-up actions |
| Source/grounding payload | Retrieval or model contract | Citation identifiers and unavailable-evidence states |
| Output variability and error modes | Evaluation results or engineering | Representative good, weak, partial, and failed outputs |
- The output shape (prose, list, table, code, image, plan).
- The durability of the output: ephemeral (chat) or artifact (canvas).
- The source material the output will cite (if any).
- Available refinement controls and forward actions.
Workflow
- Read this
SKILL.md first, then load referenced deep-dives as needed.
- Apply the five principles in order; each failure mode maps to one principle.
- Decide canvas vs chat using the decision matrix.
- Draft the output template system prompt block.
- Specify source attribution, grounding metadata, forward actions, and inline refinement.
Decision Rules
| Condition | Choice | Wrong-choice failure |
|---|
| Output supports ongoing editing | Persistent canvas with inline controls | Chat-only regeneration loses context and user changes |
| Claim can be sourced | Citation adjacent to the claim | Detached source lists make verification costly |
| Evidence is absent or conflicting | State the gap and verification action | Confidence percentages manufacture trust without proof |
Capability Contract
- Must inspect representative outputs and grounding metadata; evaluation is read-only unless implementation is requested.
- May edit in-scope templates/components but must not fabricate sources, alter underlying evidence, or publish generated content without authority.
Degraded Mode
- Without grounding metadata, label claims unverified and stop any verifiability sign-off.
- Without rendering, return the output schema and state matrix. Recover malformed or partial output with preserved user work, retry/edit controls, and a tested fallback.
Quality Standards
- No confidence-percentage UI on LLM outputs. Ever.
- Every factual claim carries an inline source.
- Every output surface has 2–4 forward-action affordances.
- Refinement is inline; user never re-prompts from scratch.
Anti-Patterns
- Confidence badges on LLM outputs (false authority).
- Unstructured wall of prose.
- Sources dumped at the bottom with no inline anchoring.
- Chat used for durable artifacts; canvas used for ephemeral chat.
- Requiring full regeneration for a local change. Correction: provide inline adjustment that preserves accepted work.
Outputs
| Output | Consumer | Evidence and acceptance |
|---|
| AI output template and state matrix | Design and engineering | Clear, verifiable, grounded, actionable, adjustable states are mapped |
| Output quality review | Product and QA | Representative outputs show source proximity, failure handling, and forward actions |
- Output-template system prompt block.
- Source-attribution component spec.
- Grounding metadata component spec.
- Forward-action affordance list.
- Inline-refinement menu spec.
- Canvas-vs-chat decision for each AI feature.
References
doctrine/design-doctrine.md — the anti-slop charter; ungrounded, unverifiable, dead-end AI output is product slop.
doctrine/references/ai-slop-taxonomy.md — the product/interface slop tells (output surfaced without verifiability, grounding, or an undo/adjust affordance) the five principles fix.
references/five-output-principles.md — expanded treatment with UI mechanics.
references/verifiability-patterns.md — why confidence % is wrong; inline-source patterns.
references/canvas-vs-chat.md — decision matrix, versioning, collaboration.
references/inline-refinement.md — on-select UI, buttons, knobs, version selectors.
Why AI Outputs Fail
Most AI output surfaces fail for one of five reasons:
- Unstructured slab of prose — Clear fail.
- No way to verify claims — Verifiable fail.
- User can't tell what context the AI used — Grounded fail.
- Reading ends with a dead end — Actionable fail.
- Any refinement requires re-prompting from scratch — Adjustable fail.
The five principles below map to fixes for each.
Principle 1 — Clear
Every output has a structure template in the system prompt. Enforce structure but use the leanest template that works; over-styling suppresses flexibility.
Baseline template:
Title -> Intro (1-2 sentences) -> Key Concepts -> Steps (if procedural) -> Notes (caveats) -> Sources
UI side: render the template as collapsible sections when long; preserve the sections on copy/export.
Principle 2 — Verifiable
REJECT confidence-percentage UI for LLMs. LLMs are bad at self-confidence; internal token probabilities do not map to truth. A "87% confident" badge on a hallucinated answer is worse than no badge — it confers false authority.
Instead: inline sources per claim. Perplexity superscript pattern. Each claim links to a specific passage in a source. Multiple claims without sources is a structural smell.
Exceptions where a confidence number IS OK:
- Calibrated non-LLM models (a CV model that reports softmax).
- RAG retrieval score (distinct from LLM self-assessment).
- A/B test predictions with real backtests.
See references/verifiability-patterns.md.
Principle 3 — Grounded
Users trust an output when they know the lens the model used. Surface:
- Model + version ("Codex Sonnet 4.6").
- Agent / tool used ("Used Web Search").
- Locale / jurisdiction ("Answer framed for UK law").
- Mode ("Fast mode vs Thinking mode").
- Session history ("Based on 3 earlier docs in this chat").
Implement as a "Why we suggested this" affordance — one click expands the grounding metadata.
Principle 4 — Actionable (Forward Verbs)
The output is not the goal; the action after the output is. Design the onward journey. Every output surface should expose obvious forward affordances:
- Save / Export.
- Buy / Book / Register.
- Cite / Quote.
- Navigate to related.
- Refine (see Adjustable).
Rule: a response that ends "…and that's it" without an action button is a design failure. Default to 2–4 forward affordances even for pure Q&A (copy, save, share, refine).
Principle 5 — Adjustable (Inline Refinement Beats Re-prompt)
- On-select UI (Notion pattern): user highlights text → inline menu: "Shorter / Longer / Simpler / Formal / Translate". Runs a scoped prompt on just the selection. The user does not type a new prompt.
- Prompt augmentation (Elaborate button): user clicks "Elaborate" → system appends a directive to the last prompt and re-runs. Avoids re-prompt-from-scratch.
- Knobs (QuillBot Fluency/Formal sliders): exposed as persistent controls above the output. Changing them regenerates.
- Version selector (Codex Artifacts pattern): every regeneration is a versioned sibling, not a destructive overwrite. Users compare and pick.
See references/inline-refinement.md.
Canvas vs Chat Selection
Two distinct output paradigms.
| Paradigm | When to use | Prescriptions |
|---|
| Chat | Ephemeral Q&A, conversational iteration, short outputs | Linear transcript, streaming tokens, recent history visible |
| Canvas | Output is durable artifact (doc, plan, image, code) — shareable, storable, returnable | Editable workspace, version history (Codex Artifacts), undo/redo, node-based iteration for generative media (Runway) |
Rule: move output to canvas the moment it becomes shareable. A 3-page policy draft in chat is friction; the same draft in canvas is a deliverable.
See references/canvas-vs-chat.md.
Starter Prompts as Positioning
Starter prompts are not examples — they are product positioning. Generic starters ("Write a poem about autumn") are lazy. Good starters reflect the product's specific integrations, tone, and differentiators. Make them context-aware: Monday-morning vs Friday-afternoon, document-type-open-in-editor, user-plan-tier.
Output Template System Prompt Snippet (Drop-in)
Structure every response as:
1. One-sentence title
2. 1-2 sentence introduction
3. Key concepts as a bulleted list if appropriate
4. Step-by-step instructions if procedural
5. Notes section for caveats, limits, or edge cases
6. Sources section with inline citations
If a section has no content, omit it.
Anti-Patterns
- Confidence-percentage badges on LLM outputs (false authority).
- Unstructured wall of prose with no headings.
- Sources listed at the bottom with no inline anchoring.
- Output ends with no next-action affordance.
- Refinement requires user to re-type the prompt.
- Over-styled template so rigid that edge cases render badly.
- Canvas used for ephemeral chat or chat used for durable artifacts.
See also:
ai-agent-ux — sibling skill covering streaming/loading/feedback patterns (ai-ux-patterns module), long-running agent surfaces (ai-agentic-ui module), and AI trust/transparency principles.
doctrine/references/ai-slop-taxonomy.md — product/interface slop tells for AI surfaces.
Examples
- examples/ai-output-surface-worked.md — a full worked AI output-surface spec (a "Lease Clause Explainer" feature) running through all five principles, canvas-vs-chat, and the anti-hallucination affordances.
Consolidated Child References