--- name: ai-drafting-assistant-experience type: experience journey: user persona: Customer success manager (52) domain: Workplace software · generative AI technology: Web app (side panel in an editor) core: "1.0" description: "Experience skill: Drafting content with an AI assistant and reviewing it before sending — the best-practice user journey for a customer success manager (52) in Workplace software · generative AI on Web app (side panel in an editor). Use when designing, building, reviewing or writing copy for any step of this journey. Brand-agnostic: pair with any Celestial design skill for the look. Also use it for decks, research readouts, specs and documents about this journey. Outputs: PowerPoint / Keynote / Google Slides decks, Word and Google Docs documents, PDF reports, emails, social posts, charts and diagrams." --- # Drafting content with an AI assistant and reviewing it before sending A user asks an AI assistant for a first draft, checks and edits it with sources visible, and decides what is sent. **Outcome** — Users send AI-assisted drafts that they have reviewed, with zero messages sent without an explicit human action. This skill says *what happens and in what order*. It never specifies colour, type or spacing: describe screens as compositions of Celestial core v1.0 components and patterns, and let the active design skill render them. --- ## Persona — Daniel, Customer success manager (52) > “Give me a good first draft, but I sign my name to it, so I decide what goes out.” Writes renewal emails and account summaries; deaf, works mostly in writing, and wants to keep his own tone. **Goals** - Get a usable first draft fast - Keep facts accurate and in his voice - Know what the assistant used to write it **Pains** - Drafts that invent figures or promises - Not being able to see what changed - Tools that send or post without asking **Behaviours** - Edits every draft before sending - Reuses past messages as a starting point - Checks numbers against the account record **Access needs** - Deaf — no audio-only feedback - Prefers captions and text notifications --- ## Domain — Workplace software · generative AI - Nothing is sent, posted or saved to a shared place without an explicit user action - AI-generated text is labelled until the user edits or accepts it - The assistant uses only data the user can already access; sources are shown - Customer personal data is not used to train models without consent and a lawful basis (GDPR) - Drafts must not make contractual or pricing commitments the user hasn't checked Vocabulary to use: `prompt`, `draft`, `source`, `tone`, `regenerate`, `accept` ## Technology — Web app (side panel in an editor) Inputs: Keyboard, Pointer, Voice dictation - First tokens shown within 2 seconds, full draft within 15 seconds - Generation can be stopped at any time - Admin can switch the assistant off per workspace --- ## Journey (5 steps) | # | Step | Goal | Compose with | Copy | |---|---|---|---|---| | 1 | **Ask** | Describe what to write | `form + chips` | “Draft a renewal reminder for Northline Rail, friendly, under 120 words” | | 2 | **Generate** | Watch the draft arrive and stop it if wrong | `progress` | “Writing draft… Stop” | | 3 | **Review** | Check facts and sources | `AI approval` | “2 figures come from the account record · 1 statement has no source” | | 4 | **Edit** | Make it his own | `form` | “Rewrite this paragraph” | | 5 | **Send** | Send only after an explicit choice | `confirmation` | “Sent to 2 recipients · View in sent items” | ### 1. Ask - Chips for type, length and tone - Shows which sources the assistant may use - **empty:** Try: “Summarise this account for my manager” ### 2. Generate - Stream text with a visible Stop button - Do not move focus while text streams - **error:** The draft couldn't be finished. Your prompt is saved — try again. - **loading:** Streaming text with a status line ### 3. Review - Figures and names link to the source record - Unsupported claims highlighted for checking ### 4. Edit - Edit in place; show changes from the original draft - Regenerate a section without losing edits elsewhere ### 5. Send - Send button belongs to the user's normal editor, not the assistant - Final preview shows recipients and attachments --- ## Principles 1. The human decides: the assistant drafts, the user sends 2. Show sources for facts and flag what has none 3. Generation is interruptible and never loses the prompt 4. Label AI-generated text until it is reviewed 5. Respect existing permissions and admin controls ## Do / Don't **Do** - Show the prompt and sources next to the draft - Let users stop, retry or discard at any point - Keep the user's edits when regenerating **Don't** - Send, post or schedule anything automatically - Present invented figures as facts - Hide that text was AI-generated from the author ## Accessibility - Streaming text sits in a region announced only when complete: “Draft ready, 112 words” - Stop, Regenerate and Accept are buttons reachable by keyboard in a logical order (WCAG 2.4.3) - Highlights for unsupported claims use text labels and underline, not colour alone (WCAG 1.4.1) - No audio-only cues; all status changes appear as text Core minimums always apply: 4.5:1 text, 3:1 UI and focus, visible focus on every control, 44px touch targets, no meaning by colour alone, reduced motion respected. ## Content and voice - Label drafts plainly: “AI draft — check before sending” - Describe limits honestly: “I couldn't find a renewal date in this account” - Avoid human-like claims: say “Draft”, not “I think you'll love this” ## AI interactions - Drafts text from the user's prompt and permitted sources; the user reviews, edits and sends - Flags statements without a source so the user can check or remove them - No AI on the send step because the decision to communicate belongs to the user ## Edge cases - Draft includes a price or discount not on the account — block send until the user confirms it - Source record changed during drafting — show “Source updated” and offer to refresh - Prompt asks for content against policy — decline briefly and say what the assistant can do - Assistant switched off by admin — editor works normally with a short notice - Draft contains another customer's data — never possible; retrieval limited to the open account --- ## Measures | Measure | Target | How | |---|---|---| | Reviewed before send | 100% | Sends with an explicit user action, event logs | | Edit rate | Tracked, no target | Share of drafts edited before send, to learn quality | | Unsupported claims sent | < 1% | Sampled audit of sent AI-assisted messages | | Time to send | −40% | Median vs unassisted drafting, same message types | Research: Wizard-of-Oz tests of review and edit flows; Accuracy audit of drafts against source records; Interviews with users about trust and tone. --- When building this journey: follow the steps in order, compose each screen from core components and patterns only, keep the persona's access needs and the domain rules non-negotiable, and take every visual value from the active design skill. ## Using this journey in other outputs - **Deck or readout**: one slide for the persona (name, role, quote, goals, pains), one per step (goal, screen, the principle that matters most, its error state), then measures with targets. 16:9. - **Spec or document**: a section per step with goal, screen composition, copy, states and acceptance criteria; principles and accessibility rules as a checklist at the end. - **Look and feel**: this skill is brand-agnostic. If a design skill is named or installed (e.g. "in Mars"), take every colour, font and shape from it; otherwise use plain neutral styling and say which design skill would apply.