👁️ Argus

The all-seeing, multi-agent, vision-verified auditor for Claude Code.

Install · Usage · Why · How

In myth, Argus Panoptes was the giant with a hundred eyes who saw everything and never fully slept. This is that, for your software: many parallel agents that test every single control for real — screenshotting before/after and looking at the result — adversarially verify each finding, and hand you a ranked, location-referenced fix list. Then /overnight keeps watch: audit → fix → verify → commit, while you sleep.

Why this exists

Most reviews check whether a feature exists. They miss the bug that actually burns users: the toggle that renders but never saves, the field that saves but never syncs to the other screen, the "Sync failed" that hides the real error. Argus checks whether every control actually works — by exercising it and looking at the screen — so you stop being the QA for your own product.

It's a method, distilled into a skill, so it works on any project — not just the one it was born on.

What makes it different

• Ground truth, seen with vision. Screenshot before/after every interaction and confirm the visible result — adb for Android, Playwright MCP for the web, run the binary for CLIs. "Renders" ≠ "works"; "saves" ≠ "syncs".
• One agent per surface, every control. Exhaustive and parallel — no button skipped.
• Fresh-eyes customer personas with zero project memory ("you're a first-timer, here's your goal") surface what's missing or confusing — plus a competitive teardown of rivals.
• Adversarial verification. Every finding gets a skeptic who tries to refute it; only confirmed defects survive. That kills false positives.
• Skill-aware. It discovers your other installed design skills and routes UI findings through them — so it polishes, not just reports.
• Universal. android-app-code · running-app · website · cli · api · mcp-server · any code.

The lenses (run all, or scope them)

functional-persistence · trust-audit (does the UI lie?) · 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.

Install

Option A — plugin marketplace (one line):

/plugin marketplace add Evil-Bane/argus
/plugin install argus

Option B — copy the skills:

git clone https://github.com/Evil-Bane/argus
cp -r argus/skills/* ~/.claude/skills/

Restart Claude Code; /customer-audit and /overnight are now available.

Usage

/customer-audit                   # full audit of the current project
/customer-audit ui                # craft + motion + a11y + tokens + delight
/customer-audit customer          # fresh-eyes personas + competitive teardown
/customer-audit https://your.app  # audit a live website via Playwright MCP
/customer-audit fix               # run, then auto-apply the confirmed fixes
/customer-audit redesign Settings # tournament: N design variants → judge → ship the winner
/overnight  make this production-ready    # autonomous audit→fix→verify→commit loop

How it works

skills/customer-audit/engine.workflow.js is a run-ready, parameterized workflow: Phase 1 fans out one agent per surface (each exercising every control + screenshotting) plus persona and competitive agents; Phase 2 adversarially verifies each finding; Phase 3 dedupes, ranks (P0–P3 with file:line / URL+selector), and emits a fix order. /overnight calls it in a loop with a durable journal so nothing is ever lost, and self-schedules its wake-ups.

Two skills ship in this plugin:

• /customer-audit — the auditor.
• /overnight — the autonomous loop that runs the auditor and fixes what it finds.

Validate it yourself

No hand-waving — reproduce the claims in a couple of minutes:

1. Install (above), then open Claude Code in any project (a small web app or repo works best).
2. Run /customer-audit ui.
3. Watch it fan out one agent per surface, exercise + screenshot each control, run an adversarial verify pass, then print a ranked P0–P3 list with file:line references.

For the autonomous loop, try /overnight fix the top 3 issues you find, build, and commit each — it keeps a durable journal and self-schedules wake-ups, so a crash or pause never loses progress.

What to check: every finding cites a concrete location; refuted findings are dropped before they reach you; UI findings are routed through any design skills you already have installed.

Contributing

Issues and PRs welcome — new lenses, new target types (desktop, game engines), and better exercise playbooks especially. Keep findings concrete and verifiable.

License

MIT © 2026 Evil-Bane