Quick Reads (Under 5 Minutes)

These days, I do my code work in Cursor with Anthropic models in a Plan-Do-Check-Act cycle.

I tried experimenting with using Haiku for some of my implementation (Do) tasks after using Sonnet for planing. The hope was to bring down cost per token on my work. But I didn’t have much luck. The results are quite understandably very poor. Despite detailed tasking from higher models, Haiku is too likely to do things like duplicate code instead of re-use it, hard code values instead of use properties, etc. etc. Too many poor patterns to enumerate. Cleaning that up exceeds any savings I’d get. I’ll wait for a model update.

In fact, I’ve started doing some of the more complicated analysis and planning with Opus. So cost per token is going up. The hope is that better analysis and planning leads to a cleaner, shorter coding cycle but I won’t have enough data and experience to establish that for a bit.

Is anyone else having luck with lower cost models for some aspects of their coding process?

More troubleshooting woes…

I am building software as generative AI emerges. The moment fills me with an awareness of the impermanence of things, Mono no aware (物の哀れ). Fascinated by the work itself. Thrilled at its possibilities. Threatened the industry discounts age and wisdom. Disgusted by our disregard for the human and environmental costs of this transition. It is joy and sadness in equal measure: complex, fragile, and with no certain future.

Even if large language models don’t improve in reasoning, there’s plenty of improvement to be gained in more precise context, additional tooling, reduced computes, increased privacy, local operation, permissively and ethically sourced training data, access, and our fluency using them.

Getting a model to build tests first is a challenge. I’ve started experimenting with this in my implementation prompt and it seems to be working.

TDD Implementation

1. ❌ DON’T test interfaces - test concrete implementations

2. ❌ DON’T use compilation errors as RED phase - use behavioral failures

3. ✅ DO create stub implementations that compile but fail behaviorally

4. ✅ DO use real components over mocks when possible

   THIS MEANS: Compilation errors are not a valid red. A red test is when an invocab does not meet the expectation. So, that would imply the project can compile and the method stubs exist but the behavior is not fully implemented.

The average US male lives to 75. The average US software developer is under 40 — tailing well off by my age. But I am not average. No human is. While much is out of my control. Some is. So, I strive to keep learning, accomplish interesting things, stay grateful, and embrace hope.

Talking to a friend, we agreed on two current challenges using claude models to troubleshoot code:

1. getting it to verify an assumption with a valid failing test before making a change to production code. “I see the problem now…” Is often (usually) not true.

2. staying calm as it declares premature victory. “✅ We fixed the problem 🎉” Is often just unwinding all the broken attempts to fix the underlying problem.

Troubleshooting code with an AI agent is like…

We’re letting hype and economic pressures compromise our professional integrity

Current research reveals that while developers feel more productive with AI tools, the data doesn’t back that up.

• 10x increase in duplicated code blocks in 2024
• Copy/pasted lines now exceed moved/refactored lines for the first time ever
• Every 25% increase in AI adoption = 7.2% decrease in delivery stability

We’re not just accumulating technical debt, we’re doing it at unprecedented scale while burning through massive computational resources at temporarily subsidized prices.

The solution isn’t sharing the industries blind faith that AI will do our jobs for us or letting our bosses tell us to just shut up and use the tools. We need to wrangle our AI practices using the disciplined approaches we already know work.

I’m proposing a human-supervised PDCA framework that might help us take more control of the technical debt generator.

Can we turn this crisis into an opportunity?

#AgileAlliance #AIcoding #TechnicalDebt #SoftwareDevelopment

My post proposing a structured plan-do-check-act prompting cycle in code generation was featured in the Agile Alliance newsletter today.