Tim Butterfield

Senior Software Engineer with 20+ years in Windows and Linux environments. Passionate about AI in software development, as explored on my blog. Tech enthusiast.

[Tim Butterfield @TimButterfield]

A powerful AI model can produce a convincing architecture before the team has found the real design question.

Stronger models help most after the workflow has shaped the problem.

#AI #SoftwareArchitecture

https://t.co/DAmEZh4RUQ

[Tim Butterfield @TimButterfield]

AI makes it cheaper to produce work that looks complete, but that does not mean the work is complete. Judgment still determines whether the result fits the architecture, preserves intent, and can be maintained.

https://t.co/rnR3Wf4ANf

#AI #SoftwareDevelopment #Architecture

[Tim Butterfield @TimButterfield]

Different AI models miss different things during review. One catches architecture drift, another catches integration issues, another catches regressions introduced during fixes.

New post:
https://t.co/yHFy3k8YC6

#AI #SoftwareEngineering

[Tim Butterfield @TimButterfield]

After a year experimenting with multi-model AI review workflows, I found the hard part was not the prompting.

It was the orchestration around the models.

Built:
- iterative review
- cross-vendor checks
- verify passes
- audit trails

https://t.co/wHDDR2Uemq

#AIEngineering #LLM

[Tim Butterfield @TimButterfield]

I wrote about running AI where it does not exist.

Built a small tool to test that idea.

Outpost: a simple bridge that forwards execution from constrained environments to machines that can run AI.

Not a framework. Just a pattern.

https://t.co/Wfp2AMpI6U

[Tim Butterfield @TimButterfield]

AI does not need to run where your code runs. It only needs a way to execute work and observe results.

I ran into this with a Windows target where Claude would not run at all. The solution was to separate execution from the AI environment.

https://t.co/t6VpDQ29b9

[Tim Butterfield @TimButterfield]

@dumay_sacha Here is an option I'm trying (hopefully, launching soon). There are the monthly plan credits per tier that expire at the end of the month. There are also add-on credits per tier (even free) that do not expire, but do get used before the plan credits. So, still pre-pay.

[Tim Butterfield @TimButterfield]

@rakyll If you're vibe coding, you don't have to worry. It will eventually delete it for you, models choice, of course. Then, you can pay it again to put it back.

Or, you can keep your repo up to date at each small point of stability to avoid some of that.

[Tim Butterfield @TimButterfield]

@ElliotSlusky @Four__ If I have a brief moment of time available, I like to be helpful.

[Tim Butterfield @TimButterfield]

@Samyak0606 If you have recursive errors, have it analyze that history and write some lessons learned to the context file. Hopefully, in the future, it can use that as instructions to do more thorough analysis and not grab the first possible fix. Also, describe the error to a different AI.

[Tim Butterfield @TimButterfield]

@ElliotSlusky You can generate a huge amount of code toward a SaaS (even after the PoC), but you still need to manually test everything that was generated to make sure it works as expected. Then, you fix all of the errors and/or gaps in functionality.

[Tim Butterfield @TimButterfield]

@Anubhavhing If you have that many left, it's past time. 😉

[Tim Butterfield @TimButterfield]

@IndieHackers ChatGPT, Claude Code, Cursor w/claude 3.7 sonnet, and sometimes Zed with other models

@PromptRefiner

[Tim Butterfield @TimButterfield]

@agidevice @cursor_ai Here are a couple of things I wish cursor did: 1) follow the rules. When the rule says perform analysis only and not make changes, don’t ignore that. 2) perform more thorough analysis and don’t automatically use the first remotely feasible match that doesn’t work, repeatedly.

[Tim Butterfield @TimButterfield]

@martin_valchev_ Continue working on @PromptRefiner. I'm hoping to complete most of the metric/costs tracking changes.

[Tim Butterfield @TimButterfield]

@iannuttall Context matters. In some countries, that could be a substantial amount.

One way to save, add this to the prompt:
"Perform analysis for a minimum change set. Do not make changes."

Verify changes first. All of the errors not created are ones you don't have to pay to fix.

[Tim Butterfield @TimButterfield]

@Harshad_57 true 😉

I usually work in dark mode, but I also test some things in light mode. For example, I needed to run some light mode tests today to verify mode handling in .svg files. After all, as you ask, people have different preferences and we want to look good for all of them.

[Tim Butterfield @TimButterfield]

@0xCaponte I was working a long contract as a senior engineer. I had a need to move and they let me go remote.

[Tim Butterfield @TimButterfield]

@gabelicona I would much rather be sleeping at that time, but my brain doesn't let me.