AI Engineering

A global progress cockpit for AI coding projects using plans.md, progress.md, and findings.md to track goals, phases, and hardpoints. Use it to reprioritize tasks, resume after context clears, and keep the big picture in focus.

Implements a Visionary/Integrator workflow between Fable and Codex to start projects, refactor code, validate plans, and delegate builds.

Write short, accurate instruction files for coding agents, including commands, conventions, and hard constraints. Use it when setting up or refreshing CLAUDE.md, AGENTS.md, or similar files.

A method for agents to persist critical state—facts, failures, conclusions—in a file that survives context compaction and session ends, with strategies for retrieval and preventing stale data.

Enforce rules deterministically in AI agents using lifecycle hooks. Automate formatting, block dangerous commands, and inject context without depending on model memory.

A guide for creating skills that trigger reliably and change an AI agent's behavior. Use when writing or debugging a skill, focusing on scoping, description writing, and progressive disclosure.

Provides guidance for creating reusable slash commands for AI agents, covering argument handling, workflow composition, and distinguishing commands from skills.

Design and dispatch subagents for context isolation, parallel work, or verification. Use when delegating tasks to a separate agent with its own context is more efficient than processing in the main context.

Provides patterns for designing reliable LLM agents with tool use, including loop termination, error handling, human checkpoints, and recognizing when an agent isn't needed.

Manage what an LLM sees by budgeting the context window, retrieving relevant information, compacting history, and pruning tool results to prevent attention degradation.

Guides the decision to fine-tune versus prompt engineering or retrieval, and provides a rigorous process for dataset construction, training with LoRA or full fine-tuning, and evaluation to achieve reliable performance improvements.

Optimize LLM feature costs by identifying where tokens are spent and applying caching, context reduction, model routing, batching, and output limits without degrading output quality.