{"id":5655,"date":"2026-04-17T08:15:21","date_gmt":"2026-04-17T08:15:21","guid":{"rendered":"https:\/\/www.aviator.co\/blog\/?p=5655"},"modified":"2026-04-17T08:50:33","modified_gmt":"2026-04-17T08:50:33","slug":"inside-runbooks-how-spec-driven-development-works","status":"publish","type":"post","link":"https:\/\/www.aviator.co\/blog\/inside-runbooks-how-spec-driven-development-works\/","title":{"rendered":"Inside Runbooks: How Spec-Driven Development Works"},"content":{"rendered":"\n
\"\"<\/figure>\n\n\n\n

Coding with AI agents is fast, but speed isn\u2019t always a good thing. In the rush to write prompts quickly, planning can easily be left behind. Agents will do what you tell them to (most of the time, at least), and before you know it, you\u2019re deep into implementation. This often results in inconsistent output, unpredictable behavior on complex codebases, and very little insight into why the agent did what it did.<\/p>\n\n\n\n

As you can see, when it comes to agentic setups, there\u2019s a clear absence of the planning phase. There are no shared specifications that both the agent and the developer can rely on. So, how should we introduce them, you may ask?<\/p>\n\n\n\n

Meet Aviator Runbooks. It uses Claude (no shade, GPT) to plan and execute tasks in a sandbox. You provide a description of what you want, review the specs it generates, and then let the agent implement them.<\/p>\n\n\n\n

The Issue of Prompting Without a Direction<\/h2>\n\n\n\n

Ah, vibe coding: throw a prompt at an AI agent, iterate, and pray something works. This is fine for greenfield projects, where you\u2019ve got a blank template and a simple goal. The agent executes the steps, you review the output, and move on without thinking twice (or looking at the billing tab).<\/p>\n\n\n\n

Brownfield enterprise codebases are a completely different matter. You\u2019re dealing with the existing architecture, implicit conventions, and years of accumulated context that no prompt (or context window) can fully capture.<\/p>\n\n\n\n

In that environment, without structure, the AI agent may need to make too many undocumented micro-decisions. This does not mean that the agent is inherently bad, though. It\u2019s just that, without structure, it\u2019s forced to make bunch of decisions related to the scope, the approach, and tradeoffs that are not visible to the dev reviewing the output. In the end, it may produce code that looks<\/strong> correct but isn\u2019t really, which is what we call an AI slop<\/a>.<\/p>\n\n\n\n

Let’s hop inside Runbooks so that I can show you a step-by-step workflow that will keep you from tearing your hair out trying to figure out why Claude made 1,337 line changes.<\/p>\n\n\n\n

\"What<\/figure>\n\n\n\n

Runbooks Workflow: Plan \u2192 Execute \u2192 Collaborate<\/h2>\n\n\n\n
\"\"<\/figure>\n\n\n\n

Inside Runbooks<\/a>, it’s like chatting with your good, old ChatGPT or Claude, but with a few differences.<\/p>\n\n\n\n

Let\u2019s see how it all works.<\/p>\n\n\n\n

\n