Flow Walkthrough Skill for Claude

Flow Walkthrough Skill

Task lists, compilers, and code review verify that *screens exist and compile*. They are structurally blind to whether the flow between screens actually works for a human with a goal — the transitions, the return paths, the dead-ends, the discoverability. This skill closes that gap.

Why this exists (the three failure classes)

Real UX bugs fail at three different layers, and no single mechanism catches all three — so this skill uses three:

| Failure class | Example | Caught by |

|---|---|---|

| Dead-end / missing path | A saved record can only be *viewed*, never reopened to edit; Done calls popToRoot() and orphans an in-progress entity | Layer 1 — static nav-graph audit (no build) |

| Reachability regression | "Run round" no longer reaches the capture screen after a refactor; a Done lands on the wrong screen | Layer 2 — automated flow driving (XCUITest) |

| Discoverability | "How do I even select a contestant?" — the only affordance is a bare row tap | Layer 3 — human checklist (a UI test taps the row and *passes*) |

> The trap: a UI test happily taps a hidden control and reports PASS. Automation proves a path *works*; only a human judges whether it is *findable*. That residue is why Layer 3 is mandatory, not optional.

Input: the <flows> from PLAN.md

/apple:plan emits a <flows> block — end-to-end journeys, each a testable script. If PLAN.md has no <flows>, derive them from the phase's <mvp-features>/views and write them back into PLAN.md first (a flow that isn't written down can't be checked). A good flow names, for every persisted entity created: the step that reopens it editable, and the single least-discoverable action.

The method

Layer 1 — Static navigation-graph audit (no build required)

Grep the navigation surface and reason over it. Do this first; it's free and catches the highest-severity class.

• Build the graph. Find entry points and edges:
• Grep: NavigationStack, navigationDestination, NavigationLink, \.sheet, \.fullScreenCover, router.push, enum Route, dismiss(), popToRoot, @Query.
• Assert reachability & return paths. For every screen: how do you get in, and does every Done/Back/dismiss land somewhere intentional? Flag any popToRoot() that discards an in-progress or just-saved entity.
• CRUD-completeness. For every @Model with a Create path, is there a Read *and* an Update/reopen path from the persistence surface (a list/history)? A list that only opens a read-only detail is a dead-end for editing — flag it.
• Entry-point sanity. Does "New X" always create a *fresh* entity with no way back to the previous one except an incomplete list? Flag create-only loops.
• Nested navigation containers. For every NavigationLink/navigationDestination destination, check whether the destination view declares its own NavigationView/NavigationStack — a pushed view must never wrap one. This renders a second nav bar under the parent's and its leading bar items displace the back button. It's a cross-file bug: each file looks correct alone (and previews fine in isolation), so single-file review never catches it — only this pairwise check does.

Output each finding as DEAD-END / NO-EDIT-PATH / ORPHANS-ENTITY / NESTED-NAV with the file:line and the missing arrow.

Layer 2 — Automated flow driving (XCUITest + per-step screenshots)

Turn each <flow> into a UI test that taps through it and screenshots every step, so the agent can *see* the transitions and assert the destinations.

• Ensure a UITest target exists. If none, add one (xcodegen: a type: bundle.ui-testing target; or xcodebuild). Keep the generated tests — they become the Phase 5 regression suite.
• Generate one test method per flow from <steps>. After each step, attach a screenshot:

   func snap(_ name: String) {
       let s = XCTAttachment(screenshot: XCUIScreen.main.screenshot())
       s.name = name; s.lifetime = .keepAlways; add(s)
   }

Launch with any needed args (e.g. a -uiTestSeed launch argument that seeds SwiftData and, in DEBUG, flips entitlement flags so gated flows are reachable without a real purchase). Assert the destination of each step (XCTAssert(app.staticTexts["Results"].waitForExistence(timeout: 2))), especially after Done/Back — that is what catches wrong-destination bugs.

• Run and extract:

   xcodebuild test -project <App>.xcodeproj -scheme <App> \
     -destination 'platform=iOS Simulator,name=<device>' \
     -resultBundlePath .planning/walkthrough/<flowId>.xcresult
   # export per-step screenshots for the agent to read:
   xcrun xcresulttool export attachments \
     --path .planning/walkthrough/<flowId>.xcresult \
     --output-path .planning/walkthrough/<flowId>/   # adapt flag to installed Xcode

(The xcresulttool attachment-export subcommand name shifts across Xcode versions — check xcrun xcresulttool --help and adapt.)

• Read the screenshots in order → a "filmstrip." Confirm each step lands where the flow says it should. A failed assertion or a wrong screen = a flow bug with visual proof.
• One unseeded pass (fresh-install reality). Every seeded run hides zero states — run at least one pass with NO seed argument that visits each top-level screen and screenshots it, in light mode. Assert each zero state renders something intentional (an empty-state view, a hint), not a blank expanse. A user's first launch is unseeded; if every automated capture is seeded, blank first-run screens ship unseen.

Layer 3 — Human discoverability pass (the irreducible residue)

For each flow, emit a short script the human runs and rates — this is the only thing that catches "I couldn't find how to do X":

Flow F1 — Run a contest (as a host)
  1. Tap "Start a Contest"        → were you sure what to tap? (1–5)
  2. Add two contestants          → was adding the *second* obvious? (1–5)
  3. Run each contestant's round  → did you know rows were tappable? (1–5)
  4. Done on results              → did you land back where you expected? (Y/N)
  Where did you hesitate? ______

Keep it to the 2–4 flows that matter; ask for a hesitation note, not just scores — the note is where the real bug hides.

Output: .planning/WALKTHROUGH.md

# Flow Walkthrough — Phase [N]

## Layer 1 — Navigation graph (static)
- 🔴 NO-EDIT-PATH: ContestHistoryView:19 opens read-only ResultsView; a saved Contest can't be reopened to edit. → route to editable setup.
- 🔴 ORPHANS-ENTITY: ResultsView "Done" calls router.popToRoot() → dumps user home, loses the contest. → dismiss() one level.

## Layer 2 — Driven flows (filmstrips in ./walkthrough/)
| Flow | Result | Note |
|------|--------|------|
| F1 create→run→results→reopen | ❌ DEAD-END at step 6 | reopen lands on read-only results |
| F2 add second contestant | ✅ reaches capture | |

## Layer 3 — Human discoverability checklist
[the per-flow scripts above, for the user to fill in]

## Verdict
[N] dead-ends (fix before phase-complete), [N] reachability failures, discoverability pending human pass.

Store screenshots under .planning/walkthrough/<flowId>/. Fix all Layer-1 dead-ends and Layer-2 failures before the phase is marked complete (same gate discipline as visual-qa).

Honest limits

• Tests can't judge feel or findability. A green Layer-2 run with an unrated Layer-3 checklist is *not* a pass.
• Seeding is required for gated/late-state flows. Provide a DEBUG -uiTestSeed path (insert sample models, set entitlement flags) so reopen/edit/results flows are reachable without manually rebuilding state each run. Remove or guard it behind the launch arg so it never affects release.
• Discoverability findings are design changes, not bugs to "fix in code" blindly — surface them, let the human decide.

Cadence & integration

• Run per flow-slice, not per-phase. Right after you build the create→edit→results loop, walk *that* loop — don't wait for the whole phase. Bugs are cheapest the minute they're introduced.
• Complements, doesn't replace, visual-qa (which is static/code-level: colors, touch targets, view states). This skill is about *flow*, that one is about *screens*.
• Feeds Phase 5. The generated XCUITests are kept as the regression suite; hand them to testing/tdd-feature/test-spec rather than re-writing.
• Pairs with a nav-graph review lens. If /apple:review gains a flow/dead-end agent, Layer 1 can run there too; this skill is the runnable, screenshot-producing counterpart.