Adam Kinney

I’m locked in! Time to prep. 5 conversations started today. Happy Friday 🫡 #getajob

I’m officially #OpenToWork 🚀

Built Microsoft Learn (200M+ MAU) → Scaled engineering org 2→200

Led Stripe Docs team through major upgrades

Spent the past year diving into AI workflows

Now ready for my next engineering leadership role. DMs open.

The precision of code + the creativity of natural language is where AI development shines.

How are you giving your AI coding agents just enough structure without killing creativity?

🛠 My 6-Step Framework:
1️⃣ Define rules once (agents.md)
2️⃣ Break work into atomic steps
3️⃣ Use code examples
4️⃣ Enforce guardrails
5️⃣ Iterate with feedback loops
6️⃣ ‼️ Review commits like any other developer

When you vibe code with AI — only using natural language — you lose that precision.

AI coding agents shine when you give them instructions + constraints.

Add 2 rules — validate inputs + round down.

Code: 2 extra lines.
English: multiple sentences, careful ordering, and qualifiers.

Programming languages pack precision into tiny syntax.
❗Well-defined type systems are key here.

In code:

const a = 2, b = 2.5;
const c = (a * b) / 4;

Short. Exact. No ambiguity.

Plain English? Long, careful sentences to avoid confusion. Word soup. 🍲

Most people think code is harder to “speak” than natural language.
I think that’s backwards. 🤔
Code is simpler — and far more precise. Here’s why 👇

11/
Happy Friday 🌞

Go burn through your Claude usage limits. 🏎️

10/
My setup now:
💬 Claude Code in VSCode for deep, conversational building
⚡ Cursor for lightning-fast manual refactors

Individually = impressive.
Together = fastest, cleanest workflow I’ve ever had.

9/
Claude might be my main driver…
…but Cursor is still the fastest, smartest editor I’ve used.

Change an enum → Cursor predicts casing, tense, properties — across the project.

I just hit Tab. ✨

8/
📚 Tangent: People worry AI will cut juniors out of the loop.

But if AI like Claude can teach while assisting, we might see AI mentorship grow talent instead of replacing it.

7/
Before Claude, my specs were thousands of lines.

Now? About thirty.

Not side-by-side pair programming — over-the-shoulder.

Sometimes I drive. Sometimes Claude does. Sometimes… it has the better idea.

6/
The impact was instant.

✅ Code quality jumped
✅ “Cancerous growth” patterns? Gone
✅ More features in weeks than I expected in a month

Preview 2 is ahead of schedule. 🚀

5/
From the first session, I noticed something different — it understood my codebase.

Not just the file I was in… but the structure, workflow, and ripple effects of changes.

4/
It’s less like prompting an AI…
More like a Slack thread with a brilliant friend who’s also a lightning-fast, world-class developer.

3/
Claude Code is built for conversation-driven coding.

Open it in the terminal → watch it think in real time. Interrupt mid-flow, steer it, and it keeps the context tight.

2/
I knew Claude existed, but never tried it.

Jimmy made it sound like something from another world… so I gave it a shot.

1/
Last week I chatted with fellow AI dev enthusiast — the Jimmy Jones.

He told me about Claude Code.

I was deep in a manual detox of Cursor output… so I was all ears. 🤠

Have you seen this kind of drift in AI-assisted code?
How do you keep the architecture from quietly falling apart?

Would love to hear what’s working for you. 👇

This isn’t an AI failure.
This is a reminder:

Working code and coherent code are not the same.

Until we get a “solutions architect” agent — one that sees across the system — it’s still my job to hold the line on code quality.

Agents are still in the mix.
Cursor is my main environment.
Codex runs small, background changes.
ChatGPT has been useful as a second-opinion architect.

But I’ve stopped pretending the spec is enough.

So I started deleting.
Merging types.
Removing duplicate logic.
Tightening the rules.

I'm now adding stricter expectations for:
- when new types are allowed
- how data flows between packages
- real code review before merge ✅

I realized I was facing a kind of “code cancer.”
It grew quietly.
Passed the tests.
Did the job.
But it wasn’t maintainable. It wasn’t healthy.

It looked like well-written code.
But it wasn’t coherent.

Each AI-generated change made sense in isolation — but without memory of the full architecture, it just kept layering more code.

Functions with nearly identical logic existed in the same file — slightly renamed.
Types that were 95% the same were defined multiple times across packages.

Singletons used… sometimes. Other times? Slightly different versions. 🤯

The code worked. It compiled.
It had types, structure, familiar patterns.

But under the hood?

It was a mess.

Last week, everything looked like it was going great.
AI agents were shipping code fast. I was reviewing specs. Progress felt smooth. 🚢

Then I started testing.
And something felt… off.

🧵

12/
What’s your AI’s drink order?

(Asking for a friend ☕)