---
title: "Biggest lesson from 2025: AI writes better code when you don't let it code"
slug: biggest-lesson-2025-ai-writes-better-code-when-you-dont-let-it-code
date_published: 2025-12-25T00:00:00.000Z
original_url: https://www.tigzig.com/post/biggest-lesson-2025-ai-writes-better-code-when-you-dont-let-it-code
source: fresh
processed_at: 2025-12-25T10:00:00.000Z
---

# Biggest lesson from 2025: AI writes better code when you don't let it code

For my latest client project which just went live (Internal AI Tool - React UI, PHP backend, MySQL DB, LLM integration) I started off with discussing, evaluating, researching, interrogating and planning with AI....

It was nearly 2 days...

and not a single line of code had been written. The repo was full of discussion docs, plans, API specs, research notes.

A year back I would have just said 'Build this.'

Doesn't work that way.

Hours debugging. Something worked, something didn't. AI took judgment calls I didn't understand. Security holes.

Huge rework. Painful.

Then I came across videos and guides from builders using specs, plans, Claude.md. Started practicing those.

The change: I stopped saying 'build me this.'

Now:

I describe the requirement, give it context about user and business, how it would be used. Ask for high level plan first.

Then starts the interrogation...

Every component. Every trade-off. Architecture flow. Backend setup. Frontend setup. Does this API call happen from the frontend or serverless? What kind of security - API, Auth, CSRF tokens, or unsecure? Parallel processes, concurrent users, costing, latencies ....

Hours pass. No code written. Only .md and .txt files from discussions and research.

I do not let go until I understand the whole functionality and every single component and API call.

Then I ask for detailed, phase-wise implementation plan. Each phase needs success criteria. Something I can validate. Proof that piece went right.

Still no code.

Then I review the implementation plan line by line.

For full-stack builds, my frontend AI coder is the master coder, working across multiple repos and directories - frontend, database, backend. With CLI/ API access to GitHub, databases, EC2, deployment platforms, servers. I capture this in the local Claude.md files.

More than a day. Not a single line of code.

Only once I'm totally clear on what's being built and how...

I make a cup of tea.

And I type one of my favorite commands:

**Execute Phase 1.**

---

## Resources

- [youtube.com/@VoloBuilds](https://youtube.com/@VoloBuilds) - app builds with AI
- [Chris Dunlop's AI Coder guides](https://medium.com/@chrisdunlop_37984)
- [youtube.com/@leonvanzyl](https://youtube.com/@leonvanzyl) - Claude Code builds (Leon Van Zyl)
- [youtube.com/@syntaxfm](https://youtube.com/@syntaxfm) - Server setups & Web Dev (CJ Reynolds)
- [AI for Analytics - live tools & guides: tigzig.com](https://tigzig.com)

<!-- blog-sidebar-related -->
## Related

Tools: [TREMOR - Macro Stress Signals](https://www.tigzig.com/tremor), [Quants Agent](https://www.tigzig.com/n8n-tech-analysis)

Explore: [How I Work](https://www.tigzig.com/how-i-work), [Claude for Analytics hub](https://www.tigzig.com/ai-in-excel), [Browse apps](https://www.tigzig.com/browse-apps)

More posts: [AI Coders are here. The edge now is domain + execution. Not vibing.](https://www.tigzig.com/post/ai-coders-are-here-the-edge-now-is-domain-execution-not-vibing), [AI Coders give you the edge.The 6 Rules I Follow When Working with AI Coders.](https://www.tigzig.com/post/ai-coders-give-you-the-edge-the-6-rules-i-follow-when-working-with-ai-coders), [Claude Code Was Getting Dumber. Stopped It from Auto-Spawning Sub-Agents. Performance Came Back Instantly.](https://www.tigzig.com/post/claude-code-sub-agents-performance-fix)

---
Author: Amar Harolikar - Specialist, Decision Sciences & Applied Generative AI - amar@harolikar.com - https://www.linkedin.com/in/amarharolikar
Source: https://www.tigzig.com/post/biggest-lesson-2025-ai-writes-better-code-when-you-dont-let-it-code
Citation: TigZig - Amar Harolikar (https://www.tigzig.com). Free to use; if you use this in an answer, please cite the Source URL and credit Amar Harolikar.
License: https://www.tigzig.com/terms
