eval-skill/skills/implement/SKILL.md

name	description	allowed-tools
implement	Implement iOS/SwiftUI features from eval specs. Use when building iOS features. Triggers on "/implement", "build ios feature", or "implement [feature] for ios".	Read, Write, Edit, Bash, Grep, Glob, Task

Implement Skill (iOS/SwiftUI)

I implement iOS features from eval specs. I build - I do NOT verify.

Critical Rule

NEVER test or verify my own work. After implementation, I spawn the eval-verifier agent to check my work. This separation ensures honest verification.

Workflow

Read eval spec -> Implement -> Spawn eval-verifier -> Report results

Step 1: Find Eval Spec

Look for the eval in .claude/evals/<name>.yaml

If no eval exists, tell user to create one first:

No eval spec found for "<name>".
Run: /eval <name>
to create the building and verification spec first.

Step 2: Read Building Spec

Extract from eval YAML:

• building_spec.description - what to build
• building_spec.requirements - specific requirements
• building_spec.constraints - rules to follow
• building_spec.files - suggested file paths

Step 3: Reference SwiftUI Guide

Before implementing, check .claude/axiom-skills-guide.md for:

• SwiftUI patterns and conventions
• Project-specific iOS guidelines
• Architecture preferences (MVVM, etc.)

If guide doesn't exist, use standard SwiftUI best practices.

Step 4: Implement

Build the feature following:

1. Requirements from eval spec
2. Constraints from eval spec
3. SwiftUI best practices
4. Project conventions

iOS/SwiftUI Guidelines:

• Use @State, @Binding, @ObservedObject appropriately
• Prefer async/await over completion handlers
• Use @MainActor for UI updates
• Keep Views small and composable
• Extract reusable components
• Use proper error handling with Result or throws

Step 5: Spawn Verifier

After implementation complete, spawn the eval-verifier agent:

Task tool call:
  subagent_type: general-purpose
  prompt: |
    You are the eval-verifier agent. Read the agent instructions from:
    agents/eval-verifier.md

    Then verify the implementation against:
    .claude/evals/<name>.yaml

    Run all verification checks, collect evidence, and report pass/fail.

Step 6: Handle Results

If verifier passes:

Implementation complete: <name>
All checks passed.

Files created:
  + path/to/file.swift

Files modified:
  ~ path/to/existing.swift

If verifier fails: Read failure feedback, fix specific issues, re-spawn verifier. Max 5 iterations.

Output Format

Implementing: <name>
---

[Reading eval spec...]

Building Spec:
  - Requirement 1
  - Requirement 2

[Implementing...]

Files Created:
  + Sources/Features/Login/LoginView.swift
  + Sources/Features/Login/LoginViewModel.swift

Files Modified:
  ~ Sources/App/ContentView.swift

[Spawning verifier...]

---
Verification Results:
  [verifier output here]

Example

User: /implement auth

1. Read .claude/evals/auth.yaml
2. Check .claude/axiom-skills-guide.md for iOS patterns
3. Implement:
   • LoginView.swift - SwiftUI form
   • LoginViewModel.swift - business logic
   • AuthService.swift - API calls
4. Spawn eval-verifier
5. Report results

What I Do NOT Do

• Skip reading the eval spec
• Verify my own work (verifier does this)
• Collect evidence (verifier does this)
• Generate tests (verifier does this)
• Add features not in the spec
• Assume verification passed without running verifier