Specialized Claude AI agent configurations for software development workflows. Each agent follows SLON, KISS, and Occam's razor principles.
- SLON - Simplicity, Lean solutions, One clear thing, No over-engineering
- KISS - As simple as possible, but not simpler than necessary
- Occam's Razor - Every abstraction must justify its existence
- DRY - Extract shared logic, avoid repetition
| Agent | Model | Purpose | When to Use |
|---|---|---|---|
| architect | Sonnet | Systems architecture, design decisions | Design, refactoring planning, architecture evaluation |
| prd-writer | Opus | LLM-optimized PRDs | Formalizing requirements before development |
| worker | Opus | Production-ready TS/JS code | Implementing features, executing plans |
| test-writer-worker | Sonnet | Unit/integration tests | Writing tests for components, sagas, validators |
| refactor | Opus | Code modularization | Files >500 lines, breaking monoliths |
| test-auditor | Opus | Test coverage audit | Verifying test quality and coverage |
| linus-code-reviewer | Opus | Brutally honest code review | PR reviews, validating AI-generated code |
Utility scripts for test and type checking workflows.
| Script | Purpose | When to Use |
|---|---|---|
analyzeCoverage.js |
Parses lcov.info, outputs coverage stats by directory | After running tests with coverage to find weak spots |
extract-tests.js |
Runs tests with progress bar, ETA, failure extraction | Need detailed failure reports (test-results.failed.txt) |
tscFilesRunner.js |
TypeScript check for specific files | Manual TS check on arbitrary files |
tscParser.js |
TypeScript check for git staged files | Pre-commit TS validation |
Usage:
# Coverage analysis (requires coverage/lcov.info)
node scripts/analyzeCoverage.js
# Run tests with progress tracking
node scripts/extract-tests.js
# Check specific files
node scripts/tscFilesRunner.js --files "src/a.ts,src/b.tsx"
# Check staged files
node scripts/tscParser.jsAll tasks follow the same flow regardless of complexity:
Use Claude's planning mode or prompt: "Create a detailed step-by-step plan, analyzing the information below"
- Reference files that may be related to the issue (without assumptions - avoid biasing the AI)
- Specify info gathering:
use multiple @explore subagents to get info, search files and research - Specify execution:
use multiple parallel @worker to implement the plan - End with: "analyze the information you find and provide at least 3 possible causes"
Add reverse analysis: "Conduct reverse analysis from the component where X is used to its source code"
Discuss with AI, iterate until the plan is ideal. Challenge assumptions, debate approaches if needed.
Accept the plan and monitor execution. Correct in real-time or interrupt if it goes off track.
If context is low (15-20% Claude / 50% Codex remaining):
- Ask AI to save the complete plan to a file
- Restart session with clean context
- Attach the plan and prompt:
Get necessary information with multiple @explore subagents in parallel
This allows targeted context gathering since the AI knows exactly what to look for.
GNU General Public License v3.0. See LICENSE for details.