AI-first development is not what you think
Everyone talks about AI-first development like it’s about generating code faster. It’s not.
The interesting question isn’t “can AI write my code?” — it’s “what changes when AI is a first-class participant in the development process?”
I’ve been building this way for a few months now. Not as an experiment — as my actual workflow. Two products shipped (Second Brain, Singapore Legal SEO), a personal site built and deployed, all with AI as a genuine collaborator. Here’s what I’ve actually learned.
The rituals we inherited
Think about how we build software. Stand-ups. Sprint planning. Jira tickets. PR reviews. These are all rituals designed for humans coordinating with humans. They assume:
- Knowledge lives in people’s heads
- Context is expensive to transfer
- Decisions need meetings
- Code review is how we maintain quality
What if none of those assumptions hold?
What actually changes
Documentation became the product
When my collaborator is an AI that starts fresh each session, I can’t rely on “you know what I mean.” I write specs, plans, and architectural decisions explicitly — not as an afterthought, but as the primary artifact.
For khairold.com, I have a .plan/ directory with four files: PLAN.md (phases and checkboxes), MEMORY.md (every decision and its rationale), DRIFT.md (spec changes), and SESSION-LOG.md (what happened each session). Every session starts by reading these files. Every session ends by updating them.
The side effect? Better architecture. When you have to write down “why are we doing it this way?” before writing code, you catch bad ideas earlier. Documentation isn’t overhead — it’s the thinking process itself.
Sessions replaced sprints
I work in focused sessions with clear handoffs. Each session reads the previous context, executes one phase, and leaves notes for the next. The “next session” might be me continuing tomorrow, or it might be starting fresh with a new context window. Either way, the handoff is clean.
My Second Brain project was built across multiple sessions, each picking up exactly where the last left off — reading the plan, checking what’s done, executing the next phase, updating everything at the end. No standup needed. No “let me get you up to speed.” The plan files are the standup.
Brain dumps replaced meetings
This one’s literal. I built a Second Brain specifically so I could dump thoughts — from my terminal or WhatsApp — and have AI extract structure, action items, entities, and decisions. No meeting required.
The capture is fast (voice note while walking), the processing is automatic (Claude extracts and files into People, Projects, Ideas, Admin), and the retrieval is proactive (daily digests surface what I’d otherwise forget).
I capture 5–10x more thoughts than I did with any note-taking app, because the friction dropped to near zero.
The real shift
AI-first development isn’t a tool change. It’s a format change.
We’re moving from human-readable processes to machine-readable ones — and discovering that machine-readable processes are often more humane. A well-structured plan file is easier for everyone to understand than tribal knowledge spread across Slack threads.
The irony is that building for AI collaboration forces you to be clearer, more structured, and more intentional than you ever were building for humans alone.
What I’m not saying
I’m not saying AI writes all my code. It doesn’t. I’m not saying it replaces thinking. It doesn’t.
What I’m saying is: when you treat AI as a participant in your workflow rather than a code generator you paste prompts into, the shape of the work itself changes. The artifacts change. The rhythms change. The things you write down and the things you keep in your head — that ratio flips.
And that flip is the interesting part. Not the code generation. The workflow transformation.