AI Coding Assistants

Most developers now build with an AI coding assistant, so inject.dart ships first-class guidance for them. Your assistant gets the annotations, the build_runner workflow, and the compile-time DI model right — without you re-explaining it every session. There are two complementary channels: a rules document and installable Agent Skills.

AI rules file

Agent Skills

For agents that support the Agent Skills format (Claude Code and others), inject.dart provides task-focused skills under skills/. Where the rules file is one big document, each skill teaches how to perform a single concrete task and is activated only when relevant. Install them with either tool:

# Node tool — installs straight from the GitHub repo
npx skills add https://github.com/ralph-bergmann/inject.dart/tree/master/packages/inject_annotation/skills --skill '*' --agent universal

# Dart tool (https://pub.dev/packages/skills) — discovers skills from your
# dependency tree and configured GitHub registries, into your IDE's skills dir
dart pub global activate skills
skills get

Together they cover every annotation in inject_annotation (@component, @module, @inject, @provides, @singleton, @asynchronous, Qualifier, @assistedInject / @assisted / @assistedFactory, @provisionListener), the runtime contracts Provider<T> and ProvisionListener<T>, and the inject_flutter view-model API.

The core rule

Whichever channel you use, the one rule that anchors all the guidance is: inject.dart is constructor injection only — “if it builds, it runs.” A good assistant never suggests service-locator patterns (get_it, provider used as a locator, or hand-rolled registries), and re-runs code generation after every annotation change:

dart run build_runner build --delete-conflicting-outputs

Validating the skills

The skills must stay true to the generator’s actual behaviour, not assumptions about it. The strongest check — for example after the generator changes — is to run each skill: have your AI assistant carry out the task each skill teaches, then prove the result compiles and generates correctly.

1. Create a temporary pure-Dart package _scratch/ — a pubspec.yaml with resolution: workspace, inject_annotation as a dependency, and build_runner + inject_generator (plus test if you exercise the testing skill) as dev-dependencies.
2. Add _scratch to the workspace: list in the root pubspec.yaml.
3. Inside _scratch/, follow each skill to perform its task — execute the skill as a user would, rather than copying snippets out of it: inject_annotation-setup-di to scaffold the component, inject_annotation-add-injectable for a service, inject_annotation-create-module to provide a configured / @asynchronous / Qualifier-ed type, then inject_annotation-add-assisted-injection, inject_annotation-inject-provider, inject_annotation-add-provision-listener, and inject_annotation-write-tests. The Flutter-only inject_annotation-flutter-view-model is exercised in a Flutter package — examples/flutter_demo already runs it and stays green.
4. Run dart pub get, then inside _scratch/: dart run build_runner build and dart analyze.
5. Confirm the build writes outputs and dart analyze reports No issues found! — i.e. following the skills produced code that compiles and generates correctly. Skim the generated *.inject.dart to confirm the wiring matches what each skill claims.
6. Delete _scratch/ and revert the pubspec.yaml edit.

A prompt you can hand your AI tool verbatim:

With the inject.dart skills installed, create a temporary pure-Dart workspace member _scratch/ and follow each skill to carry out its task (set up DI, add an injectable, create a module, assisted injection, inject a Provider, add a provision listener, write a test) — don’t copy code out of the skills, run them. Add _scratch to the root pubspec.yaml workspace: list, run dart pub get, then dart run build_runner build and dart analyze inside it. Confirm the build writes outputs, dart analyze reports “No issues found!”, and the generated code compiles and matches what each skill describes. Show me the output, then remove _scratch/ and revert the pubspec.yaml. (Validate the Flutter view-model skill against examples/flutter_demo.)

This is how the skills were validated in the first place: run each one for real, read the generated output and analyzer result, and document only what the generator actually does.