Transcription
You probably use an AI to help you code. But what if I told you that just typing vague, unspecific prompts into a chat window – what some people call “vibe-coding” – is a dead end for any serious project. It leads to poor output, lost context, and code that's impossible to maintain. The industry knows this. A new wave of structured AI development methodologies is here to fix it. But are they any good? I decided to find out.
I took one project – building a landing page for this channel – and built it three separate times. I used three different AI frameworks: the heavyweight BMAD method, GitHub’s Spek Kit, and the fast-moving Open Spec. One of these took me eight hours to build a single page. Another took just under two. So today I'm going to show you everything: the setup, the slow parts, the first parts and the final results. I'll break down how they stack up against each other. By the end of this video, you'll know exactly which framework is right for you and which one might just be a colossal waste of time. Let's get into it.
First, the project. After my last video on BMAD, this channel went absolutely viral. I now have a massive army of about 500 subscribers. With this level of fame, a simple YouTube page is just not enough. This channel deserves its own website. Building a landing page might seem simple, so I added a few challenges to make it a real test. The tech stack was Next.js, Tailwind and Shad CN UI. More importantly, I required three live API integrations. It had to use the YouTube data API to fetch my real live subscriber count. It needed a working email sign-up form connected to the MailChimp API, And it had to display my best performing videos using embedded YouTube players. I prepared a detailed specification in an input.md file describing every section. To give each framework a fair shot, I used the exact same input file and the same model for all three attempts.
Now let's see how they did, starting with the beast – BMAD. BMAD simulates an entire agile software team using different AI agents. Installation is a single command. The initial part was all about documentation. I orchestrate the different AI personas the project manager, the UX expert, and the architect to produce massive spec documents. But the real work – and the real time sink – came after. The project was broken down into four epics with 3 to 5 stories each. For every single story, I had to manually run a loop. First, call the Scrum Master to draft the story. Then hand it to the developer to write the code. Finally, pass the finished code to the QA agent for review. This entire process for one landing page took eight hours. Why so long? Because I was chained to the computer manually orchestrating every single step. Each agent would work for several minutes, but I couldn't just walk away. I had to be there to trigger the next agent in the sequence. The process went a little faster when I opened three separate terminals one for the Scrum Master, one for the developer, and one for QA. But every time I switched agents, I had to clear the context to keep them focused. And it took up to a minute for each new persona to initialize and get up to speed. It was incredibly tiresome. By the end, I was completely drained and I genuinely felt “I could have done this faster myself”. But was the result worth it? Honestly, the result was impressive. Thanks to that dedicated UX Expert agent, the design system it created was more advanced than the others. And crucially, all the integrations worked on the first try (after I fixed my own API key issue). The final website looked and felt incredibly solid. But eight hours is a tough price to pay.
Next up was GitHub’s Spek Kit. Unlike BMAD's team of agents, Spek Kit is a lightweight tool kit. It helps the developer work better with a single AI assistant. Installation involves using “uv” a Python package installer from Astral. The process is built around four simple slash commands in your AI chat. Specify. This takes the input.md as per my request, and generates a formal spec.md file Plan. This reads the spec and creates a technical plan.md file Tasks. This breaks a plan into a checklist of small tasks. Implement. Finally, this command executes the checklist and generates the code. The cool feature of Spek Kit is the constitution.md file. Here you can define high-level project rules like “Always use Test-Driven Development” or specific design principles. The AI has to follow this constitution at all times. The difference was night and day. The entire process took just under two hours. When I came back later to add the YouTube players as a new feature, the whole specify-plan-tasks-implement cycle was incredibly fast And the implementation was brilliant. Instead of loading the heavy YouTube players immediately, at first fetched static cover images from YouTube CDN. The page loaded with just these lightweight images and the play button. Only when you hit play did it load the actual YouTube player JavaScript. This gave the page a huge performance boost. The integrations worked perfectly. The AI even hit its context window limit once. But it paused, summarized its progress, and let me continue right where it left off. This felt like a tool made for developers.
Last up was Open Spec. This one feels very similar to Spec Kit. It's fast, developer focused and wants to make the specification the source of truth. The installation instructions are simple. You first install the CLI tool globally with npm. Then you navigate to your project and run “openspec init”. This sets up the project and configures your I assistant with the right instructions. The workflow is even simpler than Spek Kit’s. It's built around a proposal system For any new work, you ask your AI to create a change proposal. The AI creates a new folder for this change. Onside it generates a proposal.md, a tasks.md and “spec deltas”, which are small files showing only the proposed changes. Once you approve, you tell the AI to implement the change. When it's done, you tell the AI to archive the change, which merges the “spec deltas” into the main project specs. This was by far the fastest experience. After a few hiccups at the start, because I forgot to add my design colors and had to restart, the first implementation was stunningly fast. It took just seven minutes for the AI to generate a fully functional website, with all three integrations working perfectly. Because it was so fast, I had time for a second iteration. I took a screenshot of the first version, gave it back to the AI, and asked it to modernize the design. A few minutes later, I had animated logos and smooth scroll revealed effects.
So three functional websites. But the processes to get there were radically different. Let's break down the core philosophies head-to-head. BMAD is a top-down, process-driven system. It's philosophy is that a highly structured process that simulates a specialized team will produce the most consistent results. The process is the star of the show. Spek Kit and Open Spec are bottom-up and developer-centric. They aim to empower the individual developer, using the AI as a powerful assistant to execute a clear plan. The developer is the star. In BMAD, the human acts as a high-level orchestrator or manager. The AI agents are specialists who own their roles like product manager or QA engineer. It is human-to-AI delegation. In Spec Kit and Open Spec, the human is the implementer, actively steering a single AI-copilot through tasks. It is human-AI collaboration. BMAD operates at the macro level. It manages the entire project lifecycle from market research to QA. It brings a very high adoption overhead. It requires changing your whole process. Spec Kit and Open Spec operate at the micro level focused on a single feature or task. The adoption overhead is low. They are lightweight toolkits that add to your existing workflow. This explains everything. BMAD took eight hours because it was building an entire governance structure. Spek Kit and Open Spec took under two because they were laser-focused on implementing a feature.
So which one should you use? The deep documentation and process control of BMAD make it a powerful choice for large scale enterprise systems. Industries like finance and healthcare are still hesitant to adopt AI for coding because it has a reputation for being risky and unpredictable. BMAD's rigorous, auditable process directly addresses these fears. It's end-to-end versioning in git creates an “audit defense blueprint” where every decision is traceable. It's worth mentioning that BMAD has a groundbreaking V6 version in the works, but it's currently in alpha for early adopters only. So I decided to stick with the stable version for this test. For a solo developer, the current version is overkill, but for a large corporation that needs that level of control, BMAD is a critical tool.
For the rest of us, the choice is between Spek Kit and Open Spec. They felt very similar in practice. Both are fast, effective and produce high-quality results. As a developer, I was able to quickly find and fix any issues that came up. If I had to pick one for my daily workflow, my personal recommendation would be Spek Kit. It is maintained by a large company, GitHub, and has a very large community with over 35,000 stars. It's just being actively developed and feels like it has a stable, long term future. Open Spec is also being developed very quickly, but the corporate backing and huge community of Spek Kit give it the edge for me right now.
The most important takeaway is this structured AI development is here and it's a huge improvement over “vibe-coding”. Whether you need the control of BMAD or the speed of Spek Kit, using a methodology will make your AI work more predictable, scalable, and professional. What do you think? Have you tried any of these frameworks? Let me know your experience in the comments below. If you want to see more deep dives into AI development tools like this, make sure to subscribe to The Gray Cat. Thanks for watching and I'll see you in the next one.