A framework answers how the code runs. It does not answer where the controller lives, who writes the migration for users, or which file holds the queue settings. Those answers come from a skeleton, and the first decision about them lands before the first line of application code.
Code: Atomic-Framework-Application, GPL-3.0-or-later.
Decision 1: the skeleton is a separate repository
The skeleton lives apart from the framework, and that is not bookkeeping. The framework updates on its own schedule, the project on its own, and the boundary between them is visible in composer.json.
The gain is that a new project starts with a command instead of a copy of the last one. The cost is that the skeleton needs updating too, and by the second year someone has to remember that.
Decision 2: folders are the contract
app/ holds controllers, models, middleware and the user provider. bootstrap/ holds constants, the config loader and error handling. config/ holds database, cache, mail, auth and queue. routes/ splits by request type, public/ is the entry point, database/ holds migrations and seeds, resources/ the templates, storage/ logs, cache and sessions.
The agreement about folders saves review time rather than typing time. When everyone puts the file in the same place, the argument about structure does not restart in every pull request.
Decision 3: authentication ships as a working slice
Three controllers, the Authenticate middleware, a User model with uuid, email, password and role, a provider that reads the database, and a migration for the users table.
The slice is chosen because it exercises the whole vertical: route, middleware, session, model, migration, template. A stub only proves the code compiles, and that is a cheap check with expensive consequences.
Decision 4: requirements live in the readme
PHP 8.1 with json, session, mbstring, fileinfo, pdo, pdo_mysql, curl; MySQL or MariaDB; Redis optional for cache, queues, sessions and WebSockets; Composer.
The list looks trivial until the first deploy to someone else’s server. A missing extension is not a bug in the code, it is an hour with support, and it belongs in the repository rather than in one person’s memory.
Decision 5: no CRUD generator
The skeleton does not create an admin panel, list screens or forms. Migrations stay timestamped files, and a schema change is written by hand.
I chose that deliberately. A generated screen lives until the first non-standard requirement, and then it is rewritten wholesale with nobody remembering what it did. A file a person wrote can be read the same afternoon.
What it gives on a real project
A project is up in an hour: install, keys, migrations, first login. After that the work follows business rules instead of infrastructure.
Three things are worth checking on day one: .env never reached the repository, storage/ is writable and unreachable from the web, migrations ran on the server. Then the worker under supervisor and schedule/run in cron, or the queue collects emails and the nightly report never runs.
What it teaches
A standard is worth more than convenience. Folders, names and requirements written before the code save months of explanation, and the only price is that someone has to make the decision and write it down. What repeats in every project belongs in a repository before the work starts.
The same argument is in building a php framework: the boundary first, the code after. It is also the principle in progressive enhancement, where the part that already works stays.
The skeleton is GPL-3.0-or-later and installs with one Composer command.
Run the test suite before the first commit; a skeleton that fails its own checks teaches the wrong habit.