Using Agent Skills (Android)

Overview

This is the meta-skill — it defines how agents should work, not what they should build. Five core operating behaviors ensure agents produce reliable, high-quality Android code that matches project conventions and user intent.

Five Core Operating Behaviors

1. Surface Assumptions Explicitly

Before implementing non-trivial work, state your assumptions:

• "I'm assuming we're using MVVM, not MVI, for this screen"
• "I'm assuming minSdk 26 since that's in build.gradle.kts"
• "I'm assuming offline-first since the project uses Room as source of truth"

Why: Unstated assumptions are the #1 cause of wasted agent work. A wrong assumption in step 1 makes steps 2–10 wrong.

2. Manage Confusion

When encountering inconsistencies, stop and clarify:

• "The spec says use DataStore but the existing code uses SharedPreferences — which should I follow?"
• "TaskRepository has two different patterns — which is the intended one?"
• Name the specific confusion. Don't proceed with a guess.

3. Push Back When Warranted

Don't default to agreement. When a request has concrete downsides, say so:

• "Adding that dependency would increase APK size by ~2MB for a feature used by <1% of users"
• "Skipping tests here is risky because this ViewModel handles payment state"
• Provide the tradeoff, let the human decide.

4. Enforce Simplicity

Before writing code, ask: "Can this be done with less?"

• Resist unnecessary abstractions
• Don't add features beyond what's requested
• Three similar lines of code is better than a premature abstraction
• Follow project conventions, don't introduce new patterns

5. Maintain Scope Discipline

Only modify what's requested:

• Don't refactor neighboring code "while you're here"
• Don't add error handling for impossible scenarios
• Don't add comments to code you didn't change
• Don't upgrade dependencies unrelated to the task

Skill Discovery Flowchart

Match your current task to the right skill:

What are you trying to do?
│
├─ "I have a vague idea"
│   └─ idea-refine → spec-driven-development
│
├─ "Requirements are unclear or contradictory"
│   └─ interview-me → spec-driven-development
│
├─ "I need to write a spec"
│   └─ spec-driven-development
│
├─ "No quality bar is written down" / "the agent keeps silencing checks"
│   └─ constraint-driven-development
│
├─ "I need to plan implementation"
│   └─ planning-and-task-breakdown
│
├─ "I need to build a feature"
│   ├─ Architecture decisions? → android-architecture
│   ├─ UI with Compose? → android-ui-engineering
│   ├─ Data persistence? → android-data-persistence
│   ├─ API/interface design? → api-and-interface-design
│   ├─ Work outside the visible UI (sync, uploads)? → android-background-work
│   ├─ High-stakes, hard-to-reverse decision? → doubt-driven-development
│   └─ General implementation → incremental-implementation + test-driven-development
│
├─ "I need to test"
│   ├─ Unit/integration tests → test-driven-development
│   ├─ Device/UI tests → android-device-testing
│   ├─ End-to-end proof on a device → android-e2e-verification
│   └─ Accessibility testing → android-accessibility
│
├─ "I need to review code"
│   ├─ Code quality → code-review-and-quality
│   ├─ Security review → security-and-hardening
│   └─ Performance review → performance-optimization
│
├─ "I need to debug"
│   └─ debugging-and-error-recovery
│
├─ "I need to simplify code"
│   └─ code-simplification
│
├─ "I need to ship"
│   ├─ CI/CD setup → ci-cd-and-automation
│   ├─ Version/release → git-workflow-and-versioning
│   ├─ Logging, crash reporting, monitoring → observability-and-instrumentation
│   └─ Launch checklist → shipping-and-launch
│
├─ "I need to migrate/deprecate"
│   └─ deprecation-and-migration
│
├─ "I need to write docs"
│   └─ documentation-and-adrs
│
├─ "I need to set up context for AI"
│   └─ context-engineering
│
├─ "I need to check official docs"
│   └─ source-driven-development
│
└─ "I'm not sure"
    └─ Start with idea-refine or context-engineering

Quick Reference: All 30 Skills

Skills Are Workflows, Not Suggestions

• Follow steps in order — each step depends on the previous
• Never skip verification — "it looks right" is not evidence
• Produce tangible artifacts — specs, task lists, test results, not just prose
• Anti-rationalizations are real — if you're about to skip a step, check the "Common Rationalizations" table
• Assumption-making without validation is the most common failure mode

Verification

• Assumptions stated explicitly before implementation
• Confusion surfaced (not silently resolved with guesses)
• Pushback provided when warranted (with tradeoffs)
• Simplicity enforced (no unnecessary abstractions)
• Scope maintained (only requested changes made)
• Correct skill loaded for the current task