Turning Chaos into a Coherent Plan
The person who writes down the thing has tremendous power — Marc Andreessen
---[This is still in draft]---
After the initial honeymoon period, we have accepted that vibe coding is not sustainable to deliver production quality applications.
Realizing this, there are already a lot of movement out there in terms of finding the best way to make AI assisted development work.
Why specs matter in the age of AI
Development tooling is evolving rapidly (Claude Code, Cursor, Kiro, etc.). Coding LLMs are becoming more capable, with larger contexts and deeper understanding.
The differentiator is context engineering: structuring and feeding the right information to LLMs.
Specs as the source of truth
Specifications (specs) should be treated as first-class citizens — the true source of truth for a development project.
For far too long specs are treated as scaffolding: written once, then abandoned as iterations happen directly in code.
This used to make sense as writing elaborated specs can be considered a waste of time when we can spend it on actual code.
The fundamental shift
With AI accelerating how fast specs can be turned into working code, the focus needs to shift.
The real challenge is no longer writing code fast but crafting and maintaining detailed, living specifications.
The core questions
- What does the overall specs structure look like?
- How do we feed just enough context per prompt — not too little, not too much?
For now, let’s focus on the first question: defining the structure of specifications.
My first approach
---[To do]---
- Specs is the source of truth and need to always be updated to reflect the code
- Use spec to generate code
- If improvements are made straight to the code, specs must be updated immediately
- Do not do one-shot prompt. Break it to smaller pieces so that AI can be course corrected.
- There should be one master set of specifications with inludes business requirements specifications and the UI design
- Prompt of requirements set need to be broken down by stages and module.
The stages below is for 0 to 1, specifically for a mobile app project.
- Stage 1: Navigation & Affordance Blueprint
- Stage 2: Data layer & Mock API
Learning from Kiro
---[To do]---
Learning from Github Spec Kit
---[To do]---
Learning from Agent OS
---[To do]---
The first draft
---[To do]---