From Prompt to Harness: Mastering AI Agents, Context Engineering, and Spec‑Driven Development

Practical Tools and Methodology

The author rebuilt a macOS‑centric agent workflow that integrates shortcut launching, window tiling, terminal multiplexing, and AI‑driven coding. The core components are:

Raycast – application launcher and clipboard history (closed‑source, free).

AeroSpace – automatic window tiling and workspace management (open‑source).

Ghostty – modern terminal client recommended by AI experts (open‑source).

Yazi – TUI three‑pane file browser with preview (open‑source).

lazygit – git operations from the terminal (open‑source).

btop – real‑time system monitoring (open‑source).

Claude Code / CodeBuddy – AI‑assisted coding, hooks, and skill libraries (closed‑source, paid).

Cockpit – self‑built cross‑machine dashboard that aggregates agent status to avoid idle waiting.

tmux – persistent remote terminal sessions (open‑source).

The most frequently recommended tools are AeroSpace and Raycast because they solve common needs for workspace switching, hotkey launching, and clipboard history.

From Prompt Engineering to Context Engineering

Prompt Engineering focuses on writing effective instructions for a single turn. As tasks become multi‑step and involve long‑running agents, static prompts are insufficient. Context Engineering manages the entire information environment an agent sees—system prompts, tool definitions, external data, conversation history, and memory—so the agent receives just enough relevant context.

Even with perfect context, an agent still needs a clear specification of what the user wants. Spec‑driven Development (SDD) proposes writing a concise spec (goals, constraints, acceptance criteria) before letting the agent generate code. GitHub’s Spec Kit provides a mature workflow for this approach: https://github.com/github/spec-kit Building on SDD, Harness Engineering adds a constraint system that keeps agents from drifting into undesirable patterns. OpenAI describes this as treating agents as horses and the harness as reins, enforcing architectural rules via linters, a concise AGENTS.md index, and an automated feedback loop that turns failures into infrastructure improvements. The OpenAI article can be found at:

https://openai.com/zh-Hans-CN/index/harness-engineering/

Agent‑Powered Skills for Automation

The author uses Claude Code with a skill library and multi‑agent setup to automate daily AI news aggregation, podcast transcription, web scraping, research, and document generation. Key skills include:

ai‑news – aggregates daily AI news from >11 sources, scores items, and de‑duplicates.

podcast‑batch – batch transcribes and analyses recent podcasts (supports --days, --skip, --lang).

web‑collect – extracts and formats data from specified sites; outputs markdown, JSON, or CSV.

research – deep‑dives on a topic with multi‑source verification and structured reporting.

ai‑practices – automatically extracts reusable AI best practices from news and podcasts.

workspace‑evolve – audits the Claude workspace and suggests optimizations based on official and community best practices.

doc‑writer – generates structured documentation from outlines, including technical guides and summaries.

tool‑builder – builds small utilities with full error handling, CLI help, and automated testing.

create‑shortcut – generates macOS shortcuts via Python‑generated plist files, supporting conditions, loops, and variables.

Reflections on AI‑Native Workflows

Writing the article manually highlighted the gap between what AI can automate and the friction of tool setup, network restrictions, and cost management. Early experiments (e.g., 2025‑AI‑experienced‑OS‑dev‑study) showed that AI‑assisted coding could reduce efficiency, but rapid model improvements have turned sustained AI usage into a net positive.

Key takeaways:

Use the best available model and interact with it continuously.

Let accumulated knowledge improve future assistance—spec‑driven development, context engineering, and harness engineering together form a feedback loop that reduces rework.

In a fast‑changing AI landscape, prioritize rapid validation and integration of valuable patterns over deep static understanding.

Overall, the workflow demonstrates how agents, when combined with disciplined engineering practices, can harvest up‑to‑date best practices, automate repetitive tasks, and continuously evolve the development environment.