Let’s talk more about agentic development. Several month ago I had a thought: the reason AI often produces poor code or software is not because it’s a bad programmer, but because we, the operators, fail to explain clearly enough what we want.
🧠 We need to consider the limitations of human cognition and the common context we share. When we say “I want a button” we already imagine it quite well (e.g., the one from your company’s UI kit), while AI only sees it as the fact that you want a button.
Then I came across this video, where the speaker explains that the new way of coding isn’t about using a better programming language or a better LLM, but about writing comprehensive specifications.
🤖 The other problem is that I don’t want to write long texts to AI. It feels boring, and I’m more productive when I’m typing code instead of text. Is there anything that can help me here? Yes —
AI itself!✨ Experiment
Last week GitHub released a tool called spec-kit (short video here). Essentially, it’s a workflow composed of several prompts. It forces you to first describe your product or feature, then the technologies you want to use. From there, it generates a specification. You can edit that spec as plain text. Once you’re satisfied, it generates tasks. Only then should you ask your agent to start development.
So instead of starting with “I want a button” you start with a rich specification of your product, its ecosystem, and your expectations. This way, the LLM understands you much better.
I quickly tried it and here are the results.
What I asked it to do:
* Generate an app where you select a background and add facts about yourself. It should place them on top of the background image.
* Use
React.I didn’t edit the spec much — just added one feature. It generated something that worked, though I later had to make
~10–15 updates with prompts outside the spec workflow. Still, I think it kept the specification in context while applying those changes. I also liked how seamlessly it integrated with GitHub Copilot.Repo:
👉 https://github.com/PavelPolyakov/background-with-facts
(see the generated specifications here)
Result:
👉 https://background-with-facts.onrender.com
WDYT?