Usage Model

Related runtime/API repository: clawperator

Core Principle

Skills are a reliability layer for execution, not a substitute for agent reasoning.

The intended operating model is two-handed:

  1. Execution hand (clawperator + these skills)
  2. dispatches Android actions
  3. interacts through accessibility APIs

  4. logs/captures outputs for downstream interpretation

  5. Reasoning hand (LLM/agent)

  6. interprets outputs
  7. decides next command
  8. handles unexpected UI states
  9. manages retries/fallbacks/escalations

What Skills Guarantee

Skills aim to provide:

  • repeatable command structure
  • known selector and parsing strategies
  • stable output formatting

Skills do not guarantee:

  • that UI structure is unchanged
  • that remote config/experiments are inactive
  • that alerts/permission/update dialogs are absent
  • that app/login/account state is valid
  1. Start from a deterministic baseline state.
  2. Run skill command.
  3. Parse outputs (and screenshot path if provided).
  4. Validate expected signal presence.
  5. If mismatch: inspect UI state, adapt command, retry with controlled strategy.
  6. Return user-facing answer with confidence/fallback note if needed.

Minimal Baseline Demo (No App-Specific Skill Required)

Use Android Settings (com.android.settings) as a universal baseline probe:

  1. close_app Settings
  2. open_app Settings
  3. short settle delay
  4. snapshot_ui (ascii)
  5. capture an ADB screenshot and persist absolute file path

This gives:

  • a device-specific text snapshot (snapshot_ui) and
  • a visual snapshot (ADB screencap path)

which together are useful for multimodal LLM interpretation before moving to app-specific skills.

You can run the packaged baseline skill directly:

./skills/com.android.settings.capture-overview/scripts/capture_settings_overview.sh app.actiontask.operator.development

Anti-Patterns

  • Blindly trusting first result line.
  • Treating skill scripts as static truth forever.
  • Embedding business decisions directly in skill scripts.
  • Ignoring non-zero exits or warning-only outputs.

Practical Takeaway

These skills reduce execution friction. The LLM/agent remains responsible for correctness.