claude-tweaks-visual-review

name	claude-tweaks:visual-review
description	Use when you want to visually review a running application in the browser — inspect UI quality, walk user journeys, discover undocumented journeys, or generate creative improvement ideas. Works standalone or as a step within /claude-tweaks:review.

Interaction style: Present single decisions via the AskUserQuestion tool (options with one marked Recommended) instead of a plain-text numbered list. For multi-item decisions, render a batch table with recommended actions pre-filled, then capture the apply-all/override decision via one AskUserQuestion call. Never make more than one AskUserQuestion call per logical decision — resolve each before showing the next. End skills with a ## Next Actions block rendered via AskUserQuestion (context-specific options, one recommended), not a navigation menu.

Visual Review — Browser-Based UI Inspection

Review a running application through the browser: first impressions, persona-based interaction, structured analysis, performance vitals, and creative reimagination. Part of the workflow lifecycle:

/claude-tweaks:build → /claude-tweaks:test → /claude-tweaks:review → /claude-tweaks:wrap-up
                                                     │
                                             [ /claude-tweaks:visual-review ]
                                              (page, journey, or discover mode)

This SKILL.md is the orientation + mode resolution layer. The mechanical browser procedures (warm-up, batch walks, vitals capture, trace on failure, the Step 1-6 review flow) live in browser-review.md in this skill's directory — that file is canonical.

When to Use

• After a build — visually inspect what was built
• To walk a documented user journey and test its "should feel" expectations
• To discover and document journeys in a brownfield project
• During /claude-tweaks:review Step 6 — invoked in full mode
• After QA test runs — leverage screenshots and page inventories for enriched review
• Standalone visual inspection of any running web application

Modes

Mode	Input	What happens
page	URL or description	Review a single page or flow. Full creative framework + vitals.
journey	journey:{name}	Walk a documented journey step by step using a single batch invocation. Each step reviewed against its "should feel" / "red flags." Vitals captured per page.
discover	discover	Explore the running app to identify and document undocumented user journeys. Vitals captured per discovered page.
recommendation	--mode=recommendation (typically passed by /review Step 6 code-only)	Detect UI changes via git diff and identify affected journeys. Returns a structured recommendation (which journeys to walk, severity) without opening the browser. No agent-browser dependency.

When to use each mode

• Journey mode is the richer review — defined persona, goal, and experiential expectations at every step. Use when a documented journey exists for the affected flow.
• Page mode is for quick checks, single-page changes, or pages that aren't part of a defined journey yet.
• Discover mode is for brownfield projects that need journey coverage bootstrapped — codebase scan + browser walkthrough produces new journey files.
• Recommendation mode is for callers that need a "should we walk anything?" signal without committing to a full browser review — used by /review Step 6 in code-only mode to surface journey suggestions in the review summary.

Mode Resolution

$ARGUMENTS controls mode and target.

Standalone (invoked directly):

/claude-tweaks:visual-review http://localhost:3000           → page mode
/claude-tweaks:visual-review journey:checkout                → journey mode
/claude-tweaks:visual-review discover                        → discover mode
/claude-tweaks:visual-review                                 → page mode, auto-detect dev URL

Pipeline context (invoked by /review):

The parent skill passes:

• Mode — full (code + visual), visual, journey:{name}, or discover
• QA data — when available from a recent /claude-tweaks:test run
• Spec context — spec number or changed files for scoping

When invoked by /review in full mode, the visual review runs after code review steps complete. In standalone visual/journey/discover modes, the code review is skipped.

Step 0.5: Mode — recommendation (short-circuit)

When the resolved mode is recommendation (typically invoked as --mode=recommendation from /review Step 6 code-only), skip the browser entirely and produce a structured recommendation instead. This mode has no agent-browser dependency and no dev URL requirement.

Procedure:

1. Resolve scope — accept the changed-files list from the parent (preferred) or fall back to git diff --name-only against the base branch.
2. Cross-reference with journeys — read docs/journeys/*.md and match the changed files against each journey's files: frontmatter. A journey is affected when at least one of its files: entries appears in the changed-file set.
3. Score severity — for each affected journey:
   • high — primary flow step files (page entry, primary action component) changed
   • medium — supporting components or non-primary steps changed
   • low — peripheral files (helpers, styles only) changed
4. Return structured recommendation:

### Visual Review Recommendation (no browser)

| Journey | Severity | Changed files in journey |
|---------|----------|--------------------------|
| {journey-name} | {high\|medium\|low} | {file1}, {file2} |

> Recommended follow-up: `/claude-tweaks:visual-review journey:{top-severity-journey-name}`

If no journeys are affected, return: No documented journeys affected by the diff. Visual review not recommended.

5. Return control — skip the remaining Steps (1-4). Do not open the browser, do not run a survey, do not append a Creative Opportunities block.

Recommendation mode is read-only and browser-free — it produces a routing signal, nothing else.

Step 1: Browser Prerequisites

See _shared/browser-detection.md for the detect / install / verify procedure (daemon auto-starts on port 4848; recovery via agent-browser doctor).

Visual-review-specific behavior when agent-browser is unavailable:

Auto mode: in addition to the standard STAGED log line from the shared procedure, write staged/visual-review-skipped.md describing the skip and the install command. Append to the auto-decision log under ## /visual-review in {run-dir}/decisions.md (per _shared/auto-decision-log.md). Surface at Review Console. The review proceeds in code-only mode without further interruption.

Interactive mode: the shared procedure prompts with install / skip options. Frame the skip choice as "skip visual review — proceed with code-only review" so the user understands the impact in this skill's context. Never silently skip — always report and offer options.

Session naming, screenshot paths, and the full operation vocabulary follow the /claude-tweaks:browse skill's conventions. See browser-review.md (Prerequisites + Session naming sections) for the canonical naming and path rules used by this skill.

Step 2: Dev URL Resolution

This skill resolves the dev URL silently when possible. The canonical resolution procedure lives in dev-url-detection.md in skills/_shared/ (probes stories/servers.yml first, then applies worktree awareness (Step 2.7 — a responding port in a worktree run may be the main checkout's server, serving the wrong code), then falls back to detection heuristics and, in auto mode, starting an ephemeral worktree server on a free port). The auto-mode behavior below extends that procedure with policy-driven skip semantics.

If dev-url-detection.md cannot yield a reachable APP_URL for this checkout:

Auto mode: Run the full dev-url-detection.md Step 3 first — in a worktree (the /flow default) it auto-starts an ephemeral server on a free port (reversible, tracked, torn down at wrap-up). Only when that yields no APP_URL (no dev command found, or the server failed to come up) do you auto-skip visual review (do not ask). Append to the auto-decision log under ## /visual-review in {run-dir}/decisions.md (per _shared/auto-decision-log.md):

- STAGED {HH:MM:SS} — Step 2: no reachable dev URL for this checkout (start attempted: {yes/no}, reason: {no-dev-command | start-timeout}). Visual review skipped — code-only mode. Surface at Review Console.

Write staged/visual-review-dev-url.md capturing what was attempted. The review proceeds in code-only mode.

Interactive mode:

Call AskUserQuestion with question: "The app doesn't seem to be running at {url}. Should I:", header: "Dev server", multiSelect: false — no option is marked (Recommended); the current text has no explicit recommendation among the three:

• Option 1 — label: "Start dev server", description: "start it on a free port and continue"
• Option 2 — label: "Try different URL", description: "provide a different URL to check"
• Option 3 — label: "Wait", description: "wait while you start the dev server yourself"

In interactive mode, only start the server with the user's consent (option 1). In auto mode, starting an ephemeral worktree server is pre-authorized (it is reversible and torn down at wrap-up) — see dev-url-detection.md "Ephemeral server start".

Step 3: Run Visual Review

Run the full review procedures from browser-review.md in this skill's directory. That file is the canonical source for everything mechanical — session lifecycle, batch journey walks, annotated screenshots, vitals capture and thresholds, trace-on-failure, and the Step 0-6 review flow (Reconnaissance, Health Check, First Impressions, Use It, Analyze, Reimagine, Report & Route).

Select the procedure block matching the resolved mode:

• Page mode — read browser-review.md "Page Mode" section, then Steps 1-6.
• Journey mode — read browser-review.md "Journey Mode" section (loads the journey, assembles the batch, walks per-step, assesses the arc), then Step 6 for the report.
• Discover mode — read browser-review.md "Discover Mode" section (Phases 1-6: codebase scan → candidates → browser walkthrough → write journey files → coverage report → handoff).

Reconnaissance (Step 0) runs before the main steps in page and journey modes — read reconnaissance.md in this skill's directory for the procedure. Discover mode finds the pages itself (Phase 1 codebase scan, Phase 2 candidate selection), so per-page reconnaissance does not apply — see discover-mode.md.

Step 4: Creative Opportunities Survey

After the visual review report is assembled (per browser-review.md Step 6: Report & Route), invoke the /claude-tweaks:design wrapper's survey mode to surface ranked Creative Opportunities — recommendations for which Impeccable creative commands (bolder / delight / animate / colorize / extract / onboard / quieter / distill / overdrive) might enhance the reviewed pages.

/claude-tweaks:design survey <changed-files> --screenshots <captured-paths>

Pass:

• The file list scoped to the review (from git diff --name-only or the spec's file list).
• The annotated screenshot paths captured during review (screenshots/browse/<session>/*.png) — the wrapper analyzes each per the criteria table in command-map.md.

Handle the wrapper's return:

Return shape	Action
{result: "ok", recommendations: [...]} with non-empty list	Render the Creative Opportunities block (template below) appended to the review report.
{result: "ok", recommendations: []}	Omit the block entirely — no opportunities surfaced is a valid outcome, not a failure.
{skipped: ...}	Omit the block. Note the skip reason inline only when it would surprise the user (e.g., "Creative survey skipped — Impeccable plugin not installed").

Creative Opportunities block template

### Creative Opportunities (from /visual-review)

| Page | Observation | Suggested command |
|------|------------|-------------------|
| /pricing | Hero feels generic — pure black on white, no personality | `/impeccable:impeccable bolder pricing` |
| /empty-cart | Empty state shows only "No items" text | `/impeccable:impeccable delight empty-cart` |

> These are recommendations only. Run any command manually if you want to apply it.

When the wrapper reports suppressed > 0 in its return, append a small note below the table: > N suggestion(s) hidden — previously declined for this spec. Reset with /claude-tweaks:design reset-recommendations <spec>.

Next Actions

When invoked directly (not by a parent skill), call AskUserQuestion with question: "What's next?", header: "Next step", multiSelect: false, and:

• Option 1 — label: "Code review (Recommended)", description: "/claude-tweaks:review {spec} — full code review"
• Option 2 — label: "Walk a journey", description: "/claude-tweaks:visual-review journey:{name} — walk a specific journey"
• Option 3 — label: "Generate QA stories", description: "/claude-tweaks:stories — generate QA stories from what was reviewed"
• Option 4 — label: "Capture ideas", description: "/claude-tweaks:capture {idea} — save ideas surfaced during the review"

This is the canonical handoff block for the skill. Mode-specific Next Actions exist in discover-mode.md (post-discover variant emphasising journey walks) and browser-review.md (post-page-review variant gated by review-source signals) for situations where the standalone block doesn't fit the mode's deliverable — they render their own independent AskUserQuestion call instead of this one, never merged with it. When invoked by a parent (/review or /init), omit Next Actions — the parent handles flow control and summary.

Component-Skill Contract

This skill is a component skill — invoked by /claude-tweaks:review (Step 6) in full mode and by /claude-tweaks:init (Phase 8) for brownfield journey discovery. Parent invocation is signaled by $PIPELINE_RUN_DIR being set (the parent is running inside an active pipeline run). When invoked by a parent, omit the ## Next Actions block — the parent owns the handoff. When invoked directly by a user (no $PIPELINE_RUN_DIR), render Next Actions as shown above.

Anti-Patterns

Pattern	Why It Fails
Silently skipping when agent-browser is unavailable	Always report the missing dependency and offer options — never skip without telling the user
Skipping First Impressions in visual review	The whole point is raw reaction before structured analysis — don't make it analytical
Starting the dev server without asking in interactive mode	Interactive mode offers to start — it doesn't force it. (Auto mode is different: starting an ephemeral worktree server on a free port is pre-authorized and tracked — see dev-url-detection.md.)
Reusing a responding port without checking it serves this worktree	In a worktree run, a server on :3000 is usually the main checkout — reviewing it reports false confidence on the wrong code. Always apply dev-url-detection.md Step 2.7 (worktree awareness) before trusting a port.
Generic visual ideas ("improve the UX")	Ideas must be concrete and implementable in the current tech stack
Running visual review without a running app	The browser can't inspect what isn't served — verify the URL responds first
Describing elements by position instead of annotated overlay number	"The button on the right" is brittle; "element [3]" is precise. Always reference annotated screenshot overlays in findings
Skipping vitals capture	Performance is a first-class finding category — every reviewed page must produce LCP/CLS/INP/TTFB/FCP values
Closing the session before saving a trace on failure	Failure reports without a trace path are not actionable — trace save first, then close
Per-step agent-browser invocations during journey walks	Use batch for journey walks — one process, one session lifecycle, fewer tokens and less latency
Batching across sessions	One agent-browser batch invocation owns a single session — never mix session names
Auto-running commands suggested by the Creative Opportunities block	The block is recommendations only. The user invokes any command manually. /visual-review never executes Impeccable creative commands directly.
Rendering the Creative Opportunities block when the wrapper returned recommendations: [] or {skipped}	An empty result is a valid outcome — omit the block entirely. Surfacing "no opportunities found" as positive signal is misleading because survey is heuristic, not exhaustive.
Duplicating browser procedures in SKILL.md	browser-review.md is canonical for all mechanical browser procedures. SKILL.md is for mode resolution, auto-mode policy, and orientation — never copy procedure detail here.

Relationship to Other Skills

Skill	Relationship
/claude-tweaks:review	Invokes /visual-review in Step 6. In full mode (code + visual), visual review runs after code review. In standalone visual/journey/discover modes, /review delegates entirely to /visual-review.
/claude-tweaks:browse	/visual-review uses /browse's conventions (session naming, screenshot path, trace path) and operation vocabulary. /browse is the conventions reference; /visual-review is the review procedure. Annotated screenshots, batch walks, vitals, and trace-on-failure all follow /browse's contract.
/claude-tweaks:stories	Both skills consume dev-url-detection.md from skills/_shared/ for URL resolution. /visual-review may recommend running /stories after discovering pages.
/claude-tweaks:journeys	/visual-review (journey mode) walks journeys created by /journeys. /visual-review (discover mode) creates new journey files.
/claude-tweaks:journey-health	The deep tier falls back to /visual-review journey:{name} when auditing a journey that has no story coverage yet.
/claude-tweaks:test	QA data from /test enriches the visual review (page inventories, caveats, screenshots). Trace-on-failure convention is shared with qa-agent.
/claude-tweaks:flow	/flow invokes /review in full mode, which delegates to /visual-review for the browser portion.
/claude-tweaks:init	Detects agent-browser availability during setup. Phase 8 delegates to /visual-review discover for brownfield journey bootstrapping.
/claude-tweaks:capture	/visual-review may recommend capturing ideas surfaced during the review.
/claude-tweaks:design	After the review report is assembled, /visual-review invokes /claude-tweaks:design survey with the captured screenshot paths and renders the resulting Creative Opportunities block in the report (anchor 2 of v4.5.0's creative surfacing system). The wrapper handles its own detection (non-frontend skips); the block is omitted when the wrapper returns no recommendations.
_shared/auto-mode-contract.md	Single source of truth for auto-mode behavior — read before adding any auto-mode handling. The "Auto mode" branches in Step 1 (browser prereqs) and Step 2 (dev URL) implement the contract's auto-skip + stage-at-Review-Console pattern.
/claude-tweaks:wrap-up	verification-brief.md's Step 2.5 safety-net gate invokes /visual-review directly (same mode resolution as /review Step 6) when a testable record reaches wrap-up without a full pass already having run. Any bug found gates demo:pending the same way /review's Step 3 Routing gates PASS.
/claude-tweaks:demo	/demo's Verification Brief digest (Step 3) is sourced from /visual-review's own report — headline result + 1-3 committed screenshots. /demo's optional "Show me live" escape hatch consumes /browse's conventions directly (the same relationship /visual-review itself has with /browse), not a re-invocation of /visual-review.