TÂCHES Claude Code Resources

TÂCHES is a collection of custom resources designed to extend Claude Code's capabilities through three main components:

• Commands (27 total): Slash commands that expand into structured workflows, such as Meta-Prompting, Todo Management, Thinking Models, and Deep Analysis.
• Skills (9 total): Autonomous workflows that can research, generate, and self-heal. Examples include creating plans, MCP servers, agent skills, and subagents.
• Agents (3 total): Specialized subagents used to audit skills, slash commands, and subagent configurations for best practices and effectiveness.

The build-macos-apps skill is designed for building professional native macOS apps in Swift using SwiftUI and AppKit. It operates via a CLI-only workflow (no Xcode required) and covers the full lifecycle: building, debugging, testing, optimizing, and shipping.

Role Model:

• User: Product Owner (describes requirements, judges results).
• Claude: Developer (implements, verifies, and reports outcomes).

To start, specify your intent from the following options:

1. Build a new app
2. Debug an existing app
3. Add a feature
4. Write/run tests
5. Optimize performance
6. Ship/release
7. Something else

The build-iphone-apps skill is designed for building professional native iPhone apps using Swift, SwiftUI, and UIKit. It supports the full development lifecycle including building, debugging, testing, optimizing, and shipping.

Key Specifications:

• Environment: CLI-only (no Xcode required).
• Target: iOS 26 with iOS 18 compatibility.
• Role Model: The user acts as the Product Owner (describing requirements and judging results), while Claude acts as the Developer (implementing, verifying, and reporting outcomes).

The Pirate Bay's public JSON API is available at https://apibay.org. No authentication is required.

Rate Limiting: The API uses IP-based rate limiting. If you receive a 429 Too Many Requests error, you will be blocked for several minutes. To avoid this, batch your requests: use the detail endpoint (/t.php) to get all necessary information for a torrent in a single call rather than making multiple separate requests for metadata.

The debug-like-expert skill is a deep analysis debugging mode designed for complex issues where standard troubleshooting has failed. It follows a methodical investigation protocol involving evidence gathering, hypothesis testing, and rigorous verification.

Key Mindset:

• Scientific Rigor: Use the scientific method to identify root causes rather than applying quick fixes.
• Skepticism: Treat code you wrote with more skepticism than unfamiliar code to avoid cognitive biases about how it 'should' work.
• Verify, Don't Assume: Every hypothesis must be tested, and every fix must be validated with evidence.

The Resources-Based MCP Server Pattern is an architectural approach for building Model Context Protocol (MCP) servers that wrap large APIs (50+ operations).

In a Traditional Architecture, every operation is exposed as an individual tool. This causes massive context consumption (e.g., 15,000–30,000 tokens) because all tool definitions are sent to the LLM at the start of every conversation.

In the Resources-Based Architecture, the server exposes only a minimal set of meta-tools for discovery and execution. The detailed operation schemas are stored as MCP resources and are loaded on-demand. This can reduce context overhead by 90-98% (e.g., from 15,000 tokens down to ~300 tokens).

Ralph is an autonomous AI coding methodology that uses a continuous Bash loop to drive the Claude Code CLI. Instead of maintaining a long, error-prone conversation history, Ralph starts every iteration with a fresh context window. This prevents 'context poisoning' and forces the agent to ground its decisions in the actual files on disk.

The Core Loop:

1. The loop reads a prompt file (e.g., PROMPT.md).
2. Claude Code executes the task.
3. The agent updates the implementation plan on disk and commits changes.
4. The agent exits, and the loop restarts immediately with a clean context.

while :; do cat PROMPT.md | claude ; done

Backpressure is the use of automated validation (tests, type checks, lints, and builds) to create a self-correcting feedback loop for Ralph. Instead of Ralph generating code and hoping it works, backpressure forces Ralph to produce working code to progress.

The Feedback Loop:

1. Ralph implements a task.
2. Validation runs (e.g., npm test).
3. If validation fails, Ralph investigates and fixes the error.
4. The loop continues until validation passes.
5. Only after passing can Ralph commit and move to the next task.

AGENTS.md is a project-specific knowledge file used to evolve the intelligence of the AI agent (Ralph). It is loaded during every loop iteration alongside the prompt. Its purpose is to capture patterns, document constraints, record learnings from failures, provide build/test commands, and share context that standard prompts lack.

Key Strategy: Start Minimal Do not pre-populate the file with guessed patterns or rules you haven't observed needing. Start with an empty file or just the basic structure and add entries only when the agent fails repeatedly or when specific project context is required.

# Operational Learnings

This file contains project-specific guidance for Ralph.

## Build/Test Commands

[To be filled as needed]

## Known Patterns

[To be filled as needed]

## Constraints

[To be filled as needed]