Design AI Entry Points and Capability Onboarding

Workflow area: AX Patterns · Tier: Advanced · Hands-on time: ~45 minutes

The most common failure of shipped AI features isn't model quality — it's placement and first contact. A sparkle icon buried in a toolbar gets discovered by 4% of users; a full-screen chat takeover annoys the 96% who came to do something else; and an empty text box that says "Ask me anything" is a promise the feature cannot keep. Shape of AI calls the fix Wayfinders: patterns that signal where AI lives and what it's actually good at. Google's PAIR Guidebook adds the discipline — set expectations early, and calibrate them to real capability. This playbook combines both into an entry-point decision plus an onboarding kit you can put in Figma this week.

When to use — and when not to

• Use it when placement is still undecidedbefore engineering commits to a chat panel by default, run the entry-point decision; retrofitting placement later costs a quarter.
• Use it when discovery metrics are badthe feature shipped, adoption is single-digit, and analytics show users never open it; the fix is usually the Wayfinder layer, not the model.
• Use it when your empty state is "Ask me anything"an unscoped invitation guarantees out-of-capability prompts and first-session disappointment.
• Skip it for internal power-user toolsa team of ten analysts who requested the feature will find it; spend the effort on pb-48's failure states instead.
• Skip the full onboarding flow for a single-verb feature"Summarize this thread" needs a labeled button, not a tour.

The recipe

1. 1.

   Decide the entry-point class before any UI. For each job-to-be-done, choose one of three placements: inline (AI appears inside the existing workflow, at the moment of need — e.g., "Improve this paragraph" on selected text), dedicated surface (a panel or page users deliberately visit — right for open-ended, multi-step work), or ambient (AI acts in the background and surfaces results — right for monitoring, triage, autocomplete). The test: where does the user's context already live? AI should come to the context, not drag the context to AI.

2. 2.

   Pressure-test the choice with a structured prompt. Paste your feature description into Claude:

   Copy

   You are an AI-experience design reviewer using Shape of AI's
   Wayfinder patterns and Google PAIR's expectation-setting guidance.
   
   Feature: [ONE-PARAGRAPH DESCRIPTION: who uses it, the job-to-be-done,
   what the model actually does well and badly]
   
   Proposed entry point: [inline / dedicated surface / ambient + where]
   
   1. Argue AGAINST my proposed placement: name two user situations
      where it fails.
   2. Propose the strongest alternative placement and its trade-off.
   3. List the top 5 tasks users will TRY that the feature CANNOT do,
      given my capability description.
   4. For each, suggest how the entry point itself can prevent the
      attempt (scoped label, constrained input, absent affordance).
   Do not flatter the proposal. Be specific to this feature.
   

   Expect output like:

   Copy

   AGAINST inline: (1) First-time users never select text, so the
   affordance is invisible until taught. (2) Long-document jobs need
   multi-turn refinement, which a floating toolbar can't hold.
   ALTERNATIVE: dedicated side panel, opened from an inline trigger —
   context stays visible, conversation gets room.
   WILL TRY / CAN'T DO: full-document rewrite (>8k words) → cap shown
   in the entry label: "Improve a selection (up to ~2 pages)".
   

3. 3.

   Design capability onboarding as three sentences, not a tour. PAIR's guidance is blunt: users skip tours and remember first experiences. Write the onboarding as: one sentence of capability ("Drafts release notes from your merged PRs"), one of limitation ("It doesn't know about work outside this repo"), one of control ("You review every draft before anything is posted"). Put it where the first use happens — the empty state — not in a modal.

4. 4.

   Replace "Ask me anything" with three scoped starter prompts. Generate candidates, then cut ruthlessly:

   Copy

   Write 10 starter prompts for the empty state of this feature.
   Rules: each must be a task the feature reliably does WELL (from my
   capability list), be under 9 words, start with a verb, and produce
   a visibly impressive result on first try. Mark which 3 best teach
   the RANGE of the feature (not 3 variations of one job).
   

   Pick the three that span the range. A first prompt that succeeds teaches more than any tooltip.

5. 5.

   Assemble the Wayfinder kit in Figma. One frame per state: entry affordance in context, empty state with the three-sentence onboarding and starter chips, first-response state, and the "tried something out of scope" redirect. Annotate each with the expectation it sets. This frame set is what goes to critique — placement decisions reviewed as sketches cost minutes; reviewed in production they cost quarters.

Make it yours

• Claude Project as your placement sparring partner: load your product's IA, capability notes and this playbook's decision test into a Project; every new AI feature proposal gets step 2's adversarial review in the same context, so critiques compound.
• Figma Make for a live empty-state test: prototype the empty state with real starter chips and put it in front of five users with no explanation — count how many can say what the feature does and doesn't do after 30 seconds.
• Ambient-first products: if your entry point is ambient (suggestions appear unprompted), invert step 3 — onboarding happens at first suggestion, and the three sentences ride along with it: what triggered this, what it's based on, how to turn it off.

Quality check

• Every entry point maps to a job-to-be-donean AI surface that exists because "we needed an AI surface" fails the placement test by definition.
• The capability sentence and the starter prompts agreeif a starter prompt invites a task your limitation sentence disclaims, users will trust neither.
• Limitations are stated in user language"no context outside this repo," not "context window constraints."
• The out-of-scope redirect existsyou designed what happens when users ask for the thing it can't do, because they will, in session one.
• Discovery doesn't rely on the sparkle icon alonesomewhere, the feature announces itself at the moment its job arises.