Articles and Posts

Here are some of my learnings using large language models to help automate cleanup on a large legacy code project:

1. Prompts must be very specific to the issue type being addressed and need to be carefully crafted. This is still a 1-to-n operation as the same issue type might appear hundreds or thousands of times in the code.

2. Avoid any extraneous noise in the prompt, instructions, and context. Context (whether to add full code, parent code, dependencies) is issue type-specific.

3. Use established solutions for solved problems, e.g. ctags for code tree traversal and diff/patch for integrating changes.

4. Model selection is important. Bigger isn’t necessarily better. Larger models can be more arbitrary and cost significantly more to run (whether you pay for it or the planet does).

5. Multi-agent systems aren’t magic. Applied thoughtfully they might act as a guard against outlier variances but each interaction needs to be thoughtfully crafted and wont make quality out of nothing.

The RFG community is where practitioners and innovators of the craft of continuous improvement learn from, support, collaborate with and befriend each other.

There are perhaps many true communities in our industry but each one is a rare and priceless in helping thoughtful humans achieve purpose, embrace hope, and build mastery.

It rarely makes sense to engineer novel solutions to solved problems.

Mark Rickmeier bummed I’m not joining. Last year’s hike in Portugal was the most important and healing time I took for myself last year. I look forward to future invites! 2025???

Let’s move past prototyping and into real world software engineering. I see demonstrations of agent workflows using large parameter LLM’s at each step including solved problems like: traversing code dependencies, parsing structured outputs, and math. Incorporating less novel components may take longer but will perform at lower compute cost and with greater predictability.

Your software is unmaintainable by design.

That is, deciders have sacrificed quality through what they care about, what they reward, and their choices. Poor quality is an unintended company value.

In these situations, accelerating tools like AI code generation are more likely to increase the problem than reduce it.

To break the pattern, it’s up to leaders to champion software quality. Start by listening to the people on your teams who know what it takes to achieve it.

Prioritize reliable, performant, accessible, useable, and scalable execution by tying it to long term business outcomes. Make quality an ethical imperative in service of your customers.

Use productivity enhancements as an opportunity to invest focus, time, and money to measurably improve quality over time.

Embrace respect for coworkers. Trust but verify, of course, AND treat dignity and well-being as a human right. We will all perform better as a result.

There’s a special kind of tired when I have narrow focused on something like writing code to my endurance limit which is best remedied by either a walk outside - or - one neat glass of good whiskey.

Having options is a privilege but sometimes we have more choice than we believe. And making a choice, even a bitter one, may help you feel less trapped. With time it may help reduce regret. It has for me.

Thanks Ahmed! And congratulations.

Had no idea this was in the works. Quite a dramatic change for the tech services industry at all scales https://lnkd.in/eAW3YANE