ai-wayfinders

Use when designing AI product first-run or blank-slate UX — chat empty state, example prompts, capability discovery, templates, nudges, or follow-ups. Trigger on "users don't know what to ask", onboar

SKILL.md

GitHub

AI Wayfinders — discoverability UX

Wayfinders reduce blank-slate anxiety and teach what the product can actually do without over-claiming.

Pattern catalog: references/patterns.md. Adapted from tommyjepsen/awesome-ux-skills ai-wayfinders.

When to use

  • Chat empty state, composer placeholders, suggestion chips, command palette, capability help, first voice session.
  • Fixing "I didn't know it could X" or "it claimed Y and can't".
  • Not for: visual craft (ui-fix-observe-first), HITL confirms (ai-governors), or privacy disclosure (ai-trust-builders).

Constraints

  1. Wayfinders must track real capabilities and flags. A chip that offers a blocked or missing feature is a product bug.
  2. Voice / spoken — few short suggestions; no dense galleries.
  3. Cost honesty — suggestions that imply spend should label cost class.
  4. No dark-pattern engagement — no fake urgency; randomize only to reduce blank-slate pressure.

Surface → pattern map

Surface Prefer Avoid
Chat empty Initial CTA + 3–5 suggestions + honest cannot-do Endless gallery of unproven capabilities
Voice first start Nudge for mic permission + mode explain Auto-start without consent
Command palette Templates for frequent tasks Duplicating every nav link
Messaging channels Rare contextual follow-ups Prompt galleries in-chat
After a tool/run Follow-ups tied to this result Generic "ask me anything"
Capability / help copy Show constraints Inflated marketing bullets

Eight patterns (short)

  1. Initial CTA — scaffolded composer, not a naked box.
  2. Example Gallery — curated, actionable; dense UIs only.
  3. Suggestions — 3–6 chips; prefer contextual over static.
  4. Templates — madlibs for repeatable jobs.
  5. Nudges — one tip at the moment of need.
  6. Follow-ups — after a result, not during typing.
  7. Prompt Details — show constraints so power users learn the system.
  8. Randomize — optional shuffle of safe examples; never engagement bait.

Capability-honesty checklist

  • Mechanism exists on the live flag set (ON and OFF copy if flag-gated).
  • Medium is correct (voice tip is spoken-safe).
  • Cost class is acceptable as a default click (or labeled).
  • Failure copy distinguishes dead capability vs gate miss when possible.

Output format

  1. Blank-slate moment — which surface.
  2. Patterns — 1–3.
  3. Exact copy candidates that pass the honesty checklist.
  4. Measure — what observe-first or dogfood will verify.

Cross-references

  • ui-fix-observe-first — measure UI before declaring discoverability fixed
  • ai-trust-builders — disclosure beside onboarding
  • ai-governors — don't replace confirms with friendly chips