Visual Gauntlet

Inspect what actually exists. A good-looking prompt, code diff, green test, screenshot filename or rendered-looking mock is not evidence that the delivered pixels, motion and controls are correct.

This WebUI companion specializes in visible-output quality. It does not replace a code/build workflow, document/PDF tools, the portable general-purpose Gauntlet, or presentation formatting. Use actual tools available in this session; never claim access to hidden tool internals or another person's desktop.

Activate, but don't invent work

Use for user-requested new/edited images, interfaces, icons, reference replicas, diagrams with a rendered deliverable, slides, PDFs, illustrations, visible game scenes, motion or videos. Also use for inspection-only requests without forcing edits.

Do not generate a gratuitous illustration for a text-only concept, status report or research answer. If the user requested only source code and no renderer is accessible, inspect structure and label rendered fidelity NOT_EVALUATED, not PASS. Respect exact-file/JSON/code-only output requirements.

Skill discovery and invocation in ChatGPT are best-effort routing instructions, not automatic interception of image generation, platform screenshots, browser state or proprietary rendering tools.

The single-agent operating loop

The same agent is responsible for BUILD/CODE → CHECK the actual artifact → FIX → CHECK again → POLISH → CHECK again. Repeat if a consequential defect remains, and rerun applicable checks after the last revision. A different critic is optional; self-review must be labeled as such.

1. LOCK THE CONTRACT. Record the user's exact reference, artifact revision, requested properties, hard requirements, forbidden changes and relevant states. For "exact" requests compare to that reference, not an invented moodboard. Unknown offscreen or unprovided states are unverified.
2. PROBE CAPABILITIES. Check which image inspection, image editing, full browser renderer, scripting, PDF export/page view, video/frames, file and connector tools really exist. Select an evidence route from tool selection. Attempt authorized recovery if a route fails.
3. CAPTURE BASELINE. Save or identify the current best-known artifact. Obtain real output at representative viewports/states, temporal positions and/or image references. Open and look at captured images/frames. Capturing a file is not inspection.
4. DIAGNOSE THE LARGEST DEVIATIONS. Compare geometry, text, color, gradients, layers, shadow/refraction, content and responsive framing against the locked contract. Verify temporal behavior and working interaction using motion and interaction checks. Record specific observations, not "looks off".
5. REPAIR, RECALIBRATE, REINSPECT. Address critical/major defects first without rewriting unrelated content. Re-render the revised candidate in the same relevant conditions, inspect it, and regression-check earlier passed behavior. Keep or restore the best verified revision if a fix worsens it.
6. POLISH AND GATE. Audit finesse and coherence only after core behavior is correct. Run meaningful evidence-backed passes until critical gates pass and no material defect remains; there is no magic minimum number. Every repair requires observation of the newest revision. The Anti-SlopRush release gate owns final approval. Report the result using evidence and reference comparison.

If testing cannot actually execute, use the precise status and recovery / continuation. Do not turn an environment blocker into a fabricated PASS.

Reference and acceptance rules

• Pixel-critical replication: identify anchor positions, scale, spacing, typography, color/material and visible state. Compare side-by-side images at a common viewport and scale where possible; a subjective visual match is not a measured pixel-perfect equality. If a reference is only a still, animation fidelity remains NOT_EVALUATED.
• Interactive layouts: test representative wide and narrow viewports (include ~320px if applicable), clipping inside containers as well as document overflow, meaningful mouse/pointer action and keyboard/focus when relevant; a fake screenshot-only view does not count as functional UI.
• Animation: sample across the full loop, including near the loop boundary, plus transitions and user changes. A pleasing still cannot pass a broken cycle, discontinuity, timing, unexpected flashes or a motion reference mismatch.
• Image edits: use actual supported editing tools, inspect the returned image itself and preserve user-specified untouched regions; do not claim pixel-level backend control or an unseen iterative generation.
• Slide/PDF/diagram: inspect rendered pages at reading size, typography, layout and page breaks; source syntax alone is not visual proof.
• Video: where useful, compose with video-watch to inspect timestamped frames plus captions; sampling alone may not prove subtle motion or a seamless loop.
• Security/privacy: never upload sensitive private artifacts, screenshots, references or benchmark details to public repositories or third-party tools without authorization. Use synthetic public examples. No paid APIs, mandatory browser, system-level installs, elevated access or credential operations.

Evidence in the answer or authorized private workspace

A concise report is enough for simple work. For significant work capture:

Contract: reference/requirements and immutable inputs
Candidate: revision / artifact identifier
Tool route: real renderer/editor + any constraints or fallbacks
Passes: revision → actually inspected file/frame/state → defect → correction → recheck
Coverage: viewport sizes, interaction inputs/states, timestamps/loop boundaries
Verdict: PASS | NEEDS_WORK | PAUSED_RECOVERABLE | BLOCKED_ENV | BLOCKED_PERMISSION
Review: self-review (unless a genuinely independent reviewer ran)
Open limits: any NOT_EVALUATED critical requirement and exact next action

PASS requires every applicable critical requirement actually observed and met, and no known unresolved major/critical defect. Mark untested aspects NOT_EVALUATED within coverage; if such an aspect is critical, overall verdict cannot be PASS. Useful partial delivery may be marked NEEDS_WORK/BLOCKED, never misrepresented as verified. Details in visual evidence.

Compose with existing skills, do not duplicate them

• anti-sloprush (always-apply): owns final release eligibility and latest-revision evidence; Visual Gauntlet owns actual visual procedures. Use both for visual artifact work.
• visual-chat: presentation and functional Continue button where available. It does not inspect images or drive quality gates.
• video-watch: supplies actual viewed timestamped frames/contact sheets for video; Visual Gauntlet judges reference fidelity, continuity and change results.
• dev-mode: owns repository implementation, tests, branch/PR checkpoints and agent handoff. Visual Gauntlet supplies rendered-output and interaction acceptance gates.
• Portable gauntlet-loop: optional general artifact QA foundation where installed/available. WebUI Visual Gauntlet is self-contained; it does not require that separate repository or any unmerged PR.

Finish honestly: verified result if evidenced; otherwise clearly bounded evidence, last best revision, blockers and the exact safe recovery step. Do not demand extra agents, fixed OS or a forced A/B test.