Original: Dan Shipper (Every) · 28/01/2026
Summary
AI made us sloppy because it made us forget how to plan. While we’re on our Think Week offsite this week, we’re resurfacing Cora general manager *Kieran Klaassen**’s work on the theme of compound engineering. In this piece, Kieran argues that the best thing you can do to improve yourKey Insights
“AI made us sloppy because it made us forget how to plan.” — Discussing the negative impact of AI on planning in software development.
“One approach ships a feature. The other ships a feature and teaches the system how you think for next time.” — Comparing traditional coding to planning with AI for better long-term outcomes.
“This is compounding engineering: building systems where every unit of work makes the next one easier.” — Introducing the concept of compounding engineering to describe the benefits of planning in AI-assisted coding.
Topics
Full Article
Published: 2026-01-28
Source: https://every.to/source-code/stop-coding-and-start-planning-be0b4fd1-5898-4b09-bfda-0b00ea0004fd
.png)
While we’re on our Think Week offsite this week, we’re resurfacing Cora general manager Kieran Klaassen**’s work on the theme of compound engineering. In this piece, Kieran argues that the best thing you can do to improve your AI-assisted coding is to plan. He introduces a framework for deciding when to plan versus when to prototype, and gives a real example of how one hour of planning saved days of debugging when he wanted to turn some design plans in Figma into a product. So take that extra hour and plan. You’ll thank yourself.—Kate Lee
AI made us sloppy because it made us forget how to plan.
Planning used to be a non-negotiable part of the work: sketching screens, prototyping flows, and writing problem statements. You had to define the scope—what’s in, what’s out, what’s too ambitious, and what solves the problem. Good planning required good thinking, good writing, and collaboration between stakeholders. It was slow, but it prevented expensive mistakes.
When vibe coding emerged, planning went out the window—at first. Why spend an hour planning when you could spend five minutes building the feature? I did it, too. “Make this feature work” was my entire instruction. Sometimes it worked. Often it didn’t. When it didn’t, I’d spend three hours debugging an error that a 10-minute session—asking AI to create a clear outline of the problem and the research needed to build a solution—would have prevented. And I’d be starting from zero with each feature I shipped, instead of the AI improving with each request.
When you vibe code, you prompt, “Add email validation to the signup form,” and hope the AI takes the right route. When you plan with AI, you write: “Research how we handle validation elsewhere in the codebase, check if our email library has built-in validation, look up best practices for user-friendly error messages, then create a plan showing three approaches with tradeoffs.”
One approach ships a feature. The other ships a feature and teaches the system how you think for next time. Get this right, and the system learns from every plan. Let me show you how.
Plans teach the system—code just solves problems
I had five screens of Figma designs staring at me, and a weekend to turn these pixels into a product. We were preparing for the launch of Cora‘s email bankruptcy feature—a free service that clears users’ inbox for them without deleting anything important. Lucas Crespo and Daniel Rodrigues, Every’s designers, had turned my ugly-but-functional flow into those beautiful Figma designs: something people would want to use, with clean layouts, thoughtful interactions, and the kind of polish that sets software that delights apart from software that works. Now I had to build it. As recently as early 2025, that would have meant: Hook up the Figma MCP plugin (a tool that connects design files to code), watch it produce something vaguely related to the design but mostly ugly, then spend the weekend manually fixing it—squinting at measurements, guessing at spacing, writing HTML, refreshing the browser, noticing it’s wrong, adjusting, repeating. Days of work and frustration. This time, instead of coding all weekend, I spent one hour that saved me days. I created an AI agent with one job: Take a Figma design screenshot, analyze how to implement it, and output a detailed plan grounded in our patterns, components, and way of building.How to plan effectively: Remember the three fidelities
The first step to planning effectively…Key Takeaways
Notable Quotes
AI made us sloppy because it made us forget how to plan.Context: Discussing the negative impact of AI on planning in software development.
One approach ships a feature. The other ships a feature and teaches the system how you think for next time.Context: Comparing traditional coding to planning with AI for better long-term outcomes.
This is compounding engineering: building systems where every unit of work makes the next one easier.Context: Introducing the concept of compounding engineering to describe the benefits of planning in AI-assisted coding.
Related Topics
- [[topics/agent-native-architecture]]
- [[topics/ai-agents]]
- [[topics/prompt-engineering]]
Related Articles
Teach Your AI to Think Like a Senior Engineer
Dan Shipper (Every) · tutorial · 89% similar
Compound Engineering: Make Every Unit of Work Compound Into the Next
Dan Shipper (Every) · how-to · 82% similar
My AI Had Already Fixed the Code Before I Saw It
Dan Shipper (Every) · explanation · 81% similar
Originally published at https://every.to/source-code/stop-coding-and-start-planning-be0b4fd1-5898-4b09-bfda-0b00ea0004fd.