Onboarding Flow

A skill for designing onboarding for two very different contexts. SaaS/web onboarding and consumer/mobile onboarding fail for opposite reasons if you cross the patterns: block a SaaS user behind a linear wizard and they bounce before seeing the product; let a consumer app skip the personalization quiz and you lose the sign-up's whole reason for existing. Pick the right one before you design or evaluate anything.

This skill defines the sequence and screen types. It is not a copywriting or styling skill:

• Every word that appears on screen (headlines, button labels, quiz questions, permission copy, checklist item text) routes to the righter skill. Don't guess at wording here.
• Every structural/visual decision (spacing, components, tokens, animation, accessibility) routes to vois-patterns → vois-components → vois-tokens, in that order.
• This skill sits above both: it tells you which screens exist and in what order (or should); righter tells you what words go on them; vois tells you how they're built.

---

Two Modes

Mode 1: Evaluate an Existing Flow

1. Get the flow in front of you. Screenshots, a Figma/Mobbin link, or a plain description of the screen-by-screen sequence. If it's incomplete, say what's missing rather than guessing at unseen screens.
2. Determine the shape using the decision tree below — SaaS companion checklist or consumer linear sequence. An existing flow that uses the wrong shape for its product type is itself a finding, not something to route around.
3. Read the matching reference file and walk the flow screen by screen against every rule id in data/onboarding-rules.json for that shape, plus the five UNIV-* rules. If the flow states or implies its activation event (see Mode 2 step 2), treat any screen that doesn't move toward it as a candidate cut even when it doesn't fail a specific numbered rule.
4. Output in the Review format below. Don't rewrite copy or restyle components yourself — flag copy issues for righter and structural issues for vois-patterns, cite what should be checked there, and stop.

Mode 2: Generate a New Flow

1. What kind of product is this? Walk the decision tree below — it determines the whole shape of the flow, not just the visual style.
2. Name the activation event. Before sequencing any screens: what does this user need to accomplish for onboarding to have worked? What's the one activation event (first workflow run, first save, first invite sent, first personalized result shown), and what information is strictly necessary to reach it? Anything that doesn't materially move the user toward that event is friction, not onboarding — cut it rather than including it "just in case."
3. Pull current reference, don't rely on memory. If the mcp__Mobbin__search_flows or mcp__Mobbin__search_screens tools are available, use them to pull 3–5 real examples for the specific product category (e.g. "fintech app onboarding," "B2B analytics tool setup guide") before finalizing screens — patterns date fast, and a competitor-specific pull beats the general patterns catalogued here. If Mobbin isn't available, use references/ as the fallback baseline. Treat what you pull as evidence of current practice, not proof that a pattern is correct — weigh it against data/onboarding-rules.json, which is this skill's durable baseline. A competitor doing something doesn't override a rule here without a stated reason ("Apple does X" is not itself an argument for X).
4. Read the matching reference file for the full screen-by-screen breakdown.
5. Route copy to righter and structure to vois as you build each screen — don't batch this at the end.
6. Output in the New Flow format below.

---

Decision Tree: Which Onboarding Shape?

START: What is the user signing up for?

├─ SaaS / web product — used repeatedly at a desk, ongoing account, often paid seat-by-seat
│  └─ → read references/saas-onboarding.md                                  [SHAPE-SAAS]
│     Pattern: drop into the real product immediately, companion checklist alongside it.
│     Never a full-screen linear wizard that blocks the product.
│
└─ Consumer / mobile app — short sessions, one person, often a subscription decision up front
   └─ → read references/consumer-onboarding.md                             [SHAPE-CONSUMER]
      Pattern: linear full-screen sequence (value prop → personalize → permissions →
      account → paywall → confirmation → home). The wizard IS the experience before home.

If a product genuinely straddles both (e.g. a mobile companion app for a web SaaS product), default to the shape of the surface you're designing or evaluating this flow for — a mobile app onboarding still follows the consumer shape even if the parent product is B2B SaaS.

A second edge case: a single surface that blends both, like Notion, Figma, Slack, or a ChatGPT-style prosumer tool — used repeatedly at a desk like SaaS, but with real upfront value in a personalization choice (a workspace type, a template, a use case). Default to the SaaS shape as the base — these are still ongoing accounts used repeatedly, not short sessions — and borrow at most one consumer-style screen for the single highest-leverage personalization question, only if the product's early value genuinely depends on it. Treat CONSUMER-004's payoff rule as binding on that screen. Never let this turn into a multi-screen quiz gating access to the product — that's the SaaS shape's whole point (SAAS-001).

---

Reference Files

File	Covers	Shape ID
references/saas-onboarding.md	Setup guides, onboarding checklists, empty states, in-product companion patterns	[SHAPE-SAAS]
references/consumer-onboarding.md	Value prop screens, personalization quizzes, permission priming, paywalls, confirmation	[SHAPE-CONSUMER]

Read only the one file that matches your shape. Each is self-contained.

---

Output Formats

Review format (Mode 1)

Use this for each screen or moment reviewed. Group by screen in flow order.

---

Screen: [name/position, e.g. "Screen 3 of 6 — interest picker"]

Shape: [SaaS companion checklist | Consumer linear sequence] — flag here if the flow is using the wrong shape for its product type.

Rules checked: [list every rule id from data/onboarding-rules.json that applies to this screen type, whether it passed or failed]

Violations found: For each one:

• [rule id] (MUST | SHOULD, from data/onboarding-rules.json)
  • Observed: [what's actually on this screen in the supplied flow — point at it, don't assert without something to point at]
  • Expected: [what the rule's outcome requires]
  • Fix: [what to do instead, in one sentence]

If none, write "None found." A MUST violation is always a violation. A SHOULD violation with a stated, legitimate reason isn't one — note the reasoning in place of a fix instead of flagging it.

Route to righter: [any copy on this screen worth a pass — labels, headlines, error states — or "Nothing flagged."]

Route to vois: [any structural/spacing/component concern worth a vois-patterns/vois-components pass, or "Nothing flagged."]

After all screens: one summary line — total violations (call out MUST violations separately, since those are non-negotiable), and the single highest-leverage fix (usually a CONSUMER-004/personalization-payoff issue, a SAAS-001/blocking-wizard issue, or any MUST violation, if present).

New Flow format (Mode 2)

Use the screen sequence itself as the deliverable — a numbered list of screens in order, each with: screen purpose, primary action, what's collected or shown, and which rule ids it satisfies. Follow with the Checklist Before Handing Off below, completed, not just shown blank.

---

Universal Rules (apply regardless of shape or mode)

These held across every real flow pulled for this skill, SaaS and consumer alike. Every rule below carries a stable id in backticks — cite the id, not the bullet position, when referencing a rule from outside this file. Each is also tagged MUST or SHOULD in data/onboarding-rules.json: MUST rules address coercion, deception, or dark patterns and have no legitimate exception; SHOULD rules are strong defaults with real exceptions — deviate from them only with a stated reason, not by default.

• One primary action per screen id: UNIV-001 (SHOULD). For a screen presenting a single decision, give it exactly one primary action — if a screen has two competing CTAs for the same decision, one is winning by accident. Make the secondary action visually and verbally subordinate (text link, not a second button of equal weight). This doesn't apply to a checklist's per-item CTAs (SAAS-003 — each item is its own independent action, not competitors) or a legally required consent screen's equal-weight buttons (CONSUMER-007).
• Progress must be honest and visible id: UNIV-002 (MUST). A progress bar, step count ("Step 3 of 4"), or fraction ("3/8 completed") should reflect real remaining effort, not motivate through vagueness. Never show progress that doesn't move.
• Never trap the user id: UNIV-003 (MUST). Every screen needs a way out — back arrow, X, or skip — except the single final confirmation screen. A flow with no exit reads as coercive, not confident.
• Ask before you tell id: UNIV-004 (SHOULD). If you collect a preference (goal, interest, role), the payoff must visibly appear within one or two screens after that answer. If it won't, cut the question — asking without using it just adds friction for nothing.
• Route every word to righter, every layout decision to vois-patterns id: UNIV-005 (MUST). Don't write filler copy "for now" and don't eyeball spacing. It compounds into rework.

Structured lookup: data/onboarding-rules.json holds every tagged rule from this file and both reference files as { id, strength, condition, outcome, source_file } — useful for a quick condition/outcome check by id without opening the full file it lives in. This is also what Mode 1 walks screen-by-screen against.

Source of truth: data/onboarding-rules.json is canonical for a rule's strength, condition, and outcome. The .md files restate rules for readability and carry the worked examples (real app names, screen-by-screen breakdowns) the JSON doesn't — but if the two ever disagree, the JSON wins. scripts/check-rule-sync.mjs checks that every tagged id cited in the .md files resolves to a real entry in the JSON, the reverse, and that every rule has a valid strength; run it after editing rules in either place.

---

Checklist Before Handing Off (Mode 2)

• Shape selected (SaaS companion checklist vs. consumer linear sequence) and justified
• Activation event named, and every screen either moves toward it or was cut
• Pulled current Mobbin reference for this product's specific category, if the tool is available
• Screen sequence drafted, one primary action per screen
• Progress indicator defined and honest
• Every screen has an exit except the final confirmation
• Personalization inputs (if any) visibly reflected later in the flow
• All copy drafted through righter (not placeholder text)
• All structure/spacing/components resolved through vois-patterns → vois-components → vois-tokens

---

Maintaining This Skill

data/onboarding-rules.json is the canonical rule set (18 rules: 5 universal, 6 SaaS, 7 consumer). Every rule carries a strength of MUST or SHOULD (see Universal Rules above for what the distinction means) alongside its id, condition, outcome, and source_file. When you add, remove, or reword a rule, update it in both the .md file and the JSON, then run:

node scripts/check-rule-sync.mjs

It fails loudly if an id is tagged in markdown but missing from the JSON, listed in the JSON but never cited in markdown, tagged in a file that doesn't match the JSON's source_file, or missing a valid strength. Run it before publishing any edit to a rule.

evals/cases.json is a fixed set of worked flows (good and bad, both shapes) with the rule ids a correct review should flag — and, for the exact false positive this skill's checklist and consent patterns first shipped with, the ids it should explicitly not flag. It's a manual regression corpus, not an automated test suite: whether a flow actually violates a rule is a judgment call, not something a script can verify. Re-walk each case's scenario against the current rules after editing SKILL.md or the JSON, and confirm the same ids still come out.