Principles for developing for Chrome. These best practices center around ensuring features are implemented as efficiently as possible.
The UI that lives in the content area should be implemented with HTML technologies on desktop platforms. (Either web or native can be used for the content area UI on Android and iOS.) This HTML can be either WebUI or plain HTML and JS bundled with Chrome. It should work with no connection and lo-fi connections. The expectation is that the implementation does not contain polyfills, code that only executes for other browsers, or arcane Web development practices.
UI outside of the content area should be implemented using Views (in C++) for Windows/Linux/ChromeOS. Mac UI is done with Cocoa although work is progressing on porting Views to it as well. The expectation is that UI of this style is limited to small, transient windows with low-complexity UI. On Clank, this UI should be implemented using the Android View system.
Some features need to inspect or modify the DOM of the current tab, which can be any arbitrary website. Blink's Web* C++ API is intentionally simple and limited to guide you to use it only for what is practical. When Blink’s C++ API is impractical, use isolated world script injection. The script injected should be as lean as possible and should not contain any code that deals with other browsers (for example Closure).
Because script injection carries a runtime and memory cost, it is expected that it is driven by a user action that has an expectation of work. A good example is ‘translate’. If inspecting every page that the user might open is required, script injection is not recommended. For this case, consult with chrome-eng-review@ for an appropriate solution. For example, reader mode converted the triggering logic from script injection to native implementation.
Experimenting can be accomplished with an extension. Non-confidential features can be published on the Chrome Web Store and installed manually for early testing. Confidential extensions can be enabled only for trusted testers so they are invisible to the public. Extensions can use content scripts to do script injection and can use page actions and browser actions to provide UI outside the content area. Extensions can also use NaCL, which has full access to the Pepper API.
Component extensions (i.e. extensions automatically installed into Chrome) are also a prototyping option, but should only be used if there is something needed that a regular extension cannot do. Extensions should strive to only execute on explicit user action and never on startup. Persistent background pages should be avoided. Event pages should only be used for infrequent items like push messaging, hardware events, alarms etc…, but not common events like navigation. If you need to do something that imposes a constant load or requires updating state on each navigation, consult the extensions team; they might be able to provide a declarative API so the job can be done with low impact. Failing that, consult with chrome-eng-review@ if you need to deviate from these guidelines.