Customer audit

Skill Evil-Bane/argus/skills/customer-audit

The all-seeing, multi-agent, vision-verified auditor for Claude Code. Tests every control for real, adversarially verifies, ranks fixes. Ships /customer-audit and /overnight.

Install

npx -y skills add Evil-Bane/argus --skill customer-audit

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Universal, multi-agent, VISION-VERIFIED audit that makes any app, website, tool, CLI, API, or MCP server genuinely great. Fans out one agent per surface and exercises EVERY single control (button, field, toggle, menu, tab, swipe, long-press, drag) for real — screenshotting before/after and looking at the result — across every lens: does each control actually work + persist + sync, the "does the UI lie?" trust audit, UI/interaction craft & motion, fresh-eyes customer journeys, competitive teardown, accessibility, performance, security, microcopy, empty/error/edge states. Adversarially verifies every finding, then returns a ranked fix list (or auto-fixes). Detects and uses the available MCP servers (e.g. Playwright for websites) and routes UI findings through the user's installed design skills. Trigger when the user says "audit", "customer audit", "review every page/screen", "find everything wrong or missing", "test every button", "QA this", "make this badass/premium", "is everything working and in sync", or runs /customer-audit.

SKILL.md

7.1 KB, ~1.5k tokens by cl100k_base, as published. Nobody here has run it

/customer-audit — universal, vision-verified, multi-agent audit

Reproduce a ruthless, parallel, ground-truth audit of ANY target, the way a demanding customer + a senior engineer would together. The method is fixed; the target varies. The whole point is to catch the ONE dead control, the field that doesn't save, the screen that doesn't sync — before the user trips over it — and to make the thing feel premium. Default to maximum thoroughness.

Works on everything

android-app-code · running-app (device/emulator) · website · cli · api · mcp-server · desktop/electron · library · any code. Pick the target type, then exercise it the right way.

Modes (`$ARGUMENTS`)

/customer-audit — full audit of the current project (all lenses, all surfaces, every control).
/customer-audit ui — craft + motion + a11y + tokens + delight (routed through design skills).
/customer-audit customer — fresh-eyes personas + competitive teardown (what's missing/confusing).
/customer-audit <path|url|surface> — scope to one area / page / route / command.
/customer-audit fix — run, then AUTO-APPLY confirmed fixes (worktree-isolated, verified, committed).
/customer-audit redesign <surface> — tournament: generate N design variants, judge, ship the winner.

The seven steps

Scope. Identify the TYPE. Infer from the repo/URL; ask ONE short question only if truly ambiguous.
Inventory the environment. Detect available MCP servers (ToolSearch for "playwright", "browser", and any domain servers) and tooling (adb devices, a dev server / URL, the build/test/ run command). The right exercise engine depends on this — for a website the Playwright MCP IS how you click and see; for an app it's adb.
Discover the user's skills. List ./.claude/skills/, ~/.claude/skills/, and plugin skills. Note the design/UI skills (e.g. emil-design-eng, impeccable, frontend-design, ui-ux-pro-max, taste-skill) — READ each and apply its principles to every UI finding, and use them in redesign mode. Note deep-research (use for the competitive lens).
Enumerate surfaces AND their controls. List every screen/route/page/command, and within each, every interactive element. Fan-out is one agent per surface; each agent must hit EVERY control on its surface — none skipped (the whole point is catching the one dead one). Log anything you skip.
Configure. Lenses (default = all), a persona spread, competitors, design-skill paths, the anti-features list, a one-paragraph product context (the what, not solutions), groundTruth: true, and the MCP/tooling available.
Run the engine. Workflow({ scriptPath: "<this skill dir>/engine.workflow.js", args: {...} }). It fans out per-surface (vision-verifying every control) + fresh-eyes personas + competitive teardown, adversarially verifies every finding, dedupes, and returns the ranked report.
Deliver (or fix). Present the ranked report (location + the evidence you SAW + the fix). In fix mode, one worktree-isolated fix agent per confirmed independent finding → build/verify → commit; shared-file fixes serially. Offer to hand the backlog to /overnight.

SEE it work — vision IS the verification (the heart of this skill)

A control "rendering" proves nothing. For EVERY interactive element:

Screenshot the before. 2. Interact (tap/click/type/drag/swipe — read real element bounds from a uiautomator dump on Android or the accessibility tree / Playwright snapshot on web; never blind-tap). 3. Screenshot the after. 4. LOOK at both with vision and confirm the visible result is correct. 5. Confirm it persisted: reload / reopen and look again; on web watch the network for a 2xx + the value present; on app read it back from backend/logcat. Logs say the write fired — the screenshot says the user sees the truth.

Target	Exercise + SEE
website	Playwright MCP: `navigate` → `snapshot`/`screenshot` → `click`/`type`/`fill` → screenshot → check `network_requests` for the persisted write → reload → screenshot. Read console.
running-app	`adb`: launch → `uiautomator dump` for bounds → tap/type → `screencap` before+after → reopen to confirm it stuck → logcat for swallowed errors.
cli / api	Run/call with real + boundary + malformed inputs; assert exit/status, output, persistence (read back), error shape.
mcp-server	Enumerate its tools; call each with valid + edge args; assert schema-valid results, errors, side effects.

Lenses (engine implements all; scope with mode)

functional-persistence · trust-audit (controls that render but never persist; success shown on failed writes; optimistic updates with no rollback; stale shown as fresh) · cross-surface-sync · ui-interaction-craft · motion · accessibility · performance · security-privacy · microcopy-tone · empty-and-error-states · edge-cases · design-token-consistency · onboarding-first-run · delight · fresh-eyes-customer-journeys · competitive-gaps.

Principles that make it trustworthy

Ground truth, seen with vision. Exercise + screenshot every control. "Renders" ≠ "works"; "saves" ≠ "syncs". If you couldn't see it work, say so.
Fan out by surface; hit every control. Exhaustive + parallel.
Fresh-eyes personas carry ZERO project memory — "you are <persona>, here's your goal, first time here." Spread: first-timer, power user, impatient/break-it (double-submit, offline, weird input).
Adversarial verify. Every finding gets a skeptic who tries to refute it; keep only confirmed.
Route UI through the user's design skills; in redesign, generate 3–5 variants, judge, synthesize.
Respect anti-features (pass the "never add / removed" list to every agent).
Honest report — locations + observed evidence + concrete fixes; no inflated severity, no guesses.

See engine.workflow.js (run-ready, parameterized) for the exact agent choreography.

What ships with it: 1 file

13.2 KB alongside SKILL.md, 1 of them executable

engine.workflow.jsruns13.2 KB

Gives 0 of the 12 instructions most design frontend skills give in ~1.5k tokens

Counted across 1,169 of the 1,878 authors here whose files we hold, read 2026-08-07

Use CSS variables for color consistencyin 72 of 1169, across 23 files
Commit to one bold aesthetic direction before codingin 72 of 1169, across 27 files
Match implementation complexity to the aesthetic visionin 70 of 1169, across 20 files
Add atmospheric background effects and texturesin 57 of 1169, across 9 files
Use unexpected spatial compositions and layoutsin 56 of 1169, across 8 files
Implement real working codein 55 of 1169, across 7 files
Vary themes and aesthetics across different designsin 48 of 1169, across 7 files
Launch chromium in headless modein 47 of 1169, across 4 files
Close the browser when donein 47 of 1169, across 4 files
Run provided scripts with help flag firstin 47 of 1169, across 4 files
Wait for network idle statein 47 of 1169, across 4 files
Use descriptive selectors for elementsin 47 of 1169, across 4 files

Said here and by no other author read

exercise every interactive control exhaustively
verify every result with vision
confirm every write persisted or synced
fan out one agent per surface
route UI findings through design skills
present ranked fixes with observed evidence

Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.