Turn a web project ZIP into a repair-ready UX E2E dossier inside ChatGPT Web.
Real Chromium execution · every control mapped · source-level root cause · apply-ready patches · bilingual PDF reports
中文说明 · Download latest skill.zip · Bundled skill.zip · English sample report · 中文样例报告
A complete run does not stop at “the page looks wrong.” It produces an engineering handoff that another developer or coding agent can act on:
| You can see | You can do with it |
|---|---|
| A release decision with execution truth | Know whether the tested build is safe to ship, incomplete, or blocked |
| A numbered map of every visible control | Verify that buttons, links, fields, menus, dialogs, and responsive controls were not silently skipped |
| Expected and observed behavior chains | Follow each control from DOM event to handler, request, state mutation, rendering, navigation, and persistence |
| Layered findings with exact source excerpts | Open the faulty file and line range without repeating the investigation |
| Unified patches and regression tests | Apply the proposed repair and rerun a focused red-green verification |
| Browser screenshots, traces, logs, and manifests | Audit whether the browser run actually happened in the current ChatGPT workspace |
| Chinese and English PDF reports | Share the same evidence with product, design, QA, engineering, or external teams |
The included sample completed a fresh local Chromium run inside the ChatGPT execution environment. It accounted for 5 page states and 101 visible controls, identified a cross-region terminal-state defect with three duplicated primary progress surfaces, located the faulty source, generated one patch, and verified the repair with browser evidence plus 116/116 unit tests.
A green automated suite can still hide a broken user experience. Common examples include:
- the main panel says “completed” while the header and sidebar still say “running”;
- multiple progress bars describe the same workflow with different values;
- a button is visible but its handler, API effect, focus behavior, or final state is never tested;
- an agent reports a UI issue but cannot identify the owning source code;
- a report recommends “improving state management” without a concrete patch or acceptance test;
- historical CI artifacts are presented as if a fresh browser run happened now.
This Skill treats the application as a distributed UX state machine. It launches the uploaded project in an isolated ChatGPT environment, operates a real Chromium browser, inventories every rendered control, traces behavior to source, proposes the smallest repair in a disposable copy, reruns focused tests, and creates a decision-oriented evidence package.
flowchart LR
A[Install skill.zip in ChatGPT Web] --> B[Upload project.zip]
B --> C[Read instructions, README, tests, and source]
C --> D{Any control behavior undefined?}
D -- Yes --> E[Pause and ask the user]
D -- No --> F[Create isolated original, test, repair, and evidence copies]
F --> G[Install and start the application]
G --> H[Drive real Chromium through every reachable page and control]
H --> I[Audit cross-region state, progress, timers, loading, accessibility, and responsive behavior]
I --> J[Trace failures to exact source excerpts]
J --> K[Generate patch and regression test in repair copy]
K --> L[Run apply-check and red-green verification]
L --> M[Deliver bilingual PDFs, screenshots, traces, logs, code evidence, and patches]
The local CLI agent is only the publisher of this repository. The UX E2E audit itself runs in ChatGPT Web after the user installs the Skill and uploads a project ZIP.
- Install
skill.zipin ChatGPT Web. - Start a new chat and upload the web project as a ZIP.
- Review the result, answer only the product-behavior questions that ChatGPT cannot resolve, and hand the resulting patch/report to the implementation owner.
The long-running browser, source, patch, and reporting work is performed by ChatGPT Web.
A complete run should produce a repair handoff similar to:
ux-e2e-results/
├── report.zh-CN.pdf
├── report.en-US.pdf
├── execution-manifest.json
├── control-inventory.json
├── behavior-contracts.json
├── findings.json
├── screenshots/
│ ├── annotated/
│ ├── before/
│ └── after/
├── traces/trace.zip
├── logs/
├── source-evidence/
├── patches/
├── regression-tests/
└── run-summary.md
See What you get and Report anatomy.
Download dist/skill.zip, then upload it from the Skills interface in ChatGPT Web.
For complete results, upload the full project ZIP. It should include the lockfile, source, tests, fixtures, local startup documentation, and any safe test-only environment template.
A GitHub URL may be supplied as context, but a repository URL alone is not considered the preferred complete input. ZIP upload avoids private-repository access failures, changing branches, missing generated files, and network restrictions.
Use ChatGPT Web UX E2E Tester on the uploaded project ZIP.
Run a complete repair-oriented audit in the current isolated environment:
- install and start the application;
- use real Chromium;
- identify and visibly number every interactive control;
- stop and ask me if any material expected behavior is undefined;
- test each control's complete event/API/state/render/navigation chain;
- audit duplicated progress, loading, timers, and cross-region state consistency;
- locate faulty source with exact excerpts and line ranges;
- generate an apply-checkable unified patch and regression test in a disposable repair copy;
- rerun focused browser and existing tests;
- deliver Chinese and English PDF reports plus the evidence and patch bundle.
More prompt variants are in examples/prompts.
A complete run may take 30-90 minutes. One observed ChatGPT Web run using a non-Pro model took approximately 60 minutes. This is expected: the Skill installs dependencies, starts services, inventories controls, performs real browser actions, captures traces and screenshots, diagnoses source, prepares patches, reruns tests, and renders reports.
A long run is not automatically a stalled run. The execution manifest and final report must distinguish completed, blocked, artifact-review, and static-only modes.
Designed for web applications including React, Next.js, Vue, static HTML, and full-stack projects that can be started in the isolated environment.
The Skill does not bypass live CAPTCHA, OTP, OAuth, payment, anti-abuse, or third-party security controls. External services that cannot be safely replaced are reported as blocked rather than falsely passed.
skills/chatgpt-web-ux-e2e-tester/ canonical Skill source
dist/skill.zip installable ChatGPT Skill
examples/reports/ bilingual sample reports
docs/images/ effect-first report previews
tools/ validation and release builders
scripts/publish-with-gh.sh safe GitHub CLI publisher
.github/workflows/ repository validation and release automation
- What you get
- How it works
- ChatGPT Web quickstart
- Project ZIP requirements
- Runtime expectations
- Limitations
- Security and privacy
- Publishing with a local Agent CLI
The initial public release is a beta. The sample demonstrates the intended evidence standard, but project compatibility still depends on the packages, services, browser policies, test fixtures, and external dependencies available in the ChatGPT execution environment.
The repository is licensed under Apache-2.0. Adapted third-party patterns and notices are documented in THIRD_PARTY_NOTICES.md.