markng.co.uk

What It Is

demos-not-memos is a developer tool built by Mark Ng around the acceptance workflow for AI coding agents. Its core premise is that when agents such as Claude Code or Codex claim that an implementation is “done,” developers still have to read diffs, trace logic, and manually click through the UI to confirm whether the feature actually meets expectations. This tool tries to close that validation loop by having the agent directly “demo what it built,” using browser interactions with synchronized narration.

Core Capabilities

The article shows that the project is a TypeScript DSL built on top of Playwright and distributed as the npm package demos-not-memos. In the example, users can create a NarratedDemo, set the baseUrl, voice, and video output path, then call start, narrate, Playwright locator actions, and finish to generate a narrated browser demo. The author has also packaged it as a Claude Code skill/plugin called narrated-demo.

In terms of supported tools, the text explicitly mentions Claude Code, and an update says it can be used with OpenAI Codex. Since the npm package is just TypeScript and Playwright, with no Claude-specific dependency, it should theoretically be callable by other AI coding tools as well. However, the article does not describe support for non-Web UIs, mobile apps, backend APIs, or CLI scenarios.

Pricing and Ecosystem

The collected content does not provide information about pricing, payment methods, licensing, whether it is open source, self-hosting, or enterprise support. It appears as both an npm package and a Claude Code plugin, suggesting an installation and integration path that is friendly to developers. However, it lacks complete documentation, version compatibility details, a security model, and maintenance commitments, so it still feels more like an early-stage personal tool.

Pros, Cons, and Who It’s For

Its main strength is that it addresses a real pain point in AI-assisted coding: reducing the disconnect where “the agent writes the code, but humans do the QA,” so code review can focus more on quality rather than whether the intent was understood. Being based on Playwright also makes it easier for developers to understand and extend.

The downsides are limited disclosure, with unknown maturity, stability, and long-term maintenance. Its value is also mainly concentrated on validating features with a browser-based UI.

It is best suited to individual developers and small teams that frequently use Claude Code or Codex to build Web prototypes or frontend features, especially for generating acceptance demos or quickly spotting misunderstandings in requirements. For enterprise teams with strict compliance needs or SLA requirements, the currently available information is insufficient.

Access from China

The article does not provide information about availability in mainland China, mirrors, payments, or network accessibility, so this remains unknown. Potential alternatives or complementary tools include Playwright, Cypress, Selenium, as well as the testing/demo capabilities within Claude Code and OpenAI Codex themselves.