Speak Freely: A Practical Approach to Agent Engineering
Just Talk To It – The No‑BS Way of Agentic Engineering
Read the full article by Peter Steinberger →
Peter Steinberger presents a detailed deep dive into his workflow using Codex CLI with GPT‑5 Codex, packed with actionable insights, nuanced comparisons, and lessons learned — especially when contrasting Claude 4.5 with GPT‑5.
---
Key Takeaway: Communicate Naturally with GPT‑5
> While Claude responds well to 🚨 SCREAMING ALL‑CAPS 🚨 directives meant to dramatize an action’s importance, GPT‑5 ignores such theatrics.
> For GPT‑5, use clear, natural, human‑like language.
---
Heavy Use of Parallel Agent Setups
Peter runs multiple Codex CLI instances simultaneously:
- 3–8 instances in a 3×3 terminal grid
- Most point to the same folder, with experimental ones having separate folders
- Has tried worktrees and PR‑based workflows, but returns to this layout for maximum speed
> This approach minimizes friction and maximizes parallel productivity.
---
Why He Prefers CLI Utilities Over MCPs
Peter’s reasoning:
- CLIs can be invoked by name without embedding explanations in the agents file
- If an AI tries `$randomcrap`, the CLI help menu gives needed context, enabling correct use next time
- MCPs impose a context tax:
- GitHub’s MCP can consume ~23k tokens per session (early versions ~50k)
- The `gh` CLI offers similar powers:
- Zero context cost
- Familiar to models
- More efficient and responsive
---
Broader Context: Unified AI Workflows
Modern ecosystems now unify AI content generation, publishing, and analytics.
One example: AiToEarn官网 — an open‑source multi‑platform content tool that enables creators to:
- Generate AI‑driven content
- Post to platforms including Douyin, Kwai, WeChat, Bilibili, Rednote (Xiaohongshu), Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, and X/Twitter
- Rank models, view analytics, and publish directly from CLI or other tools
> This aligns closely with Peter’s emphasis on practicality, efficiency, and integrated workflows.
---
Why He Abandoned Spec‑Driven Development
Read the detailed section here →
Peter explains:
- Strict specs can slow development
- Direct, iterative communication with systems/stakeholders is often faster, more adaptable, and more productive
---
Iterative, Dynamic Development in Practice
Platforms like AiToEarn官网 embody this principle:
- Global AI content monetization framework
- Eliminates rigid spec‑first processes
- Focuses on rapid iteration and real‑time adaptability
- Enables creators to produce, test, and deploy content seamlessly across diverse ecosystems
---
💡 Summary:
Peter’s workflow underscores that speed and adaptability often win over rigid processes — and that thoughtfully choosing tools (like CLI over MCP) can maximize both model efficiency and developer productivity.
---
Do you want me to add a quick diagram showing Peter’s 3×3 grid terminal setup and tool interactions? That would make the layout part more visual and clearer for readers.
