AI development is changing fast enough that we need to reconsider more than yesterday’s failures. We also need to reconsider yesterday’s successful methods.
Anthropic recently announced that it removed more than 80 percent of the Claude Code system prompt for its latest models without a measurable loss in coding performance. OpenAI’s guidance and newer models appear to be moving in a similar direction.
That means the detailed prompts, repeated rules, exhaustive examples, and large context packages that helped older models may sometimes be getting in the way of newer ones.
In Roll with the Flow, I look at why AI programmers need to keep retesting both what failed and what worked:
https://askgoodquestions.dev/guides/roll-with-the-flow
In the companion field note, Watch the Tide, I relate that lesson to my own experience developing PageSnip and reconsidering the future of PromptStacker:
https://askgoodquestions.dev/field-notes/watch-the-tide
Experience still matters, but it cannot become a fixed set of procedures. The models keep changing, and we have to keep changing how we work with them.
I was recently looking into something related to keeping a lean memory file. This is to avoid having instructions that conflict with newer models.
As these models evolve, what used to hold you back, newer models can now handle with few or no instructions. The interesting thing about this is that it’s not a difference of years, just a difference of months.
Check this video. Not related to this specific post, but more tied to your series about AI
Thanks for sharing this.
He had some very interesting observations, especially about the bottleneck moving from writing code to specification, supervision, and verification, and the danger of losing the tribal knowledge that experienced developers carry in their heads.
I used PageSnip to extract and lightly clean the video transcript, so I’m attaching it as a PDF for anyone who would rather read it or keep it for reference. That is also a useful example of something PageSnip can do: turn a video into a clean, searchable document that is easy to save and share.
It fits the “Roll with the Flow” discussion very well. AI is not simply making the old development workflow faster.
It’s changing where the real work has to happen.
What 6 Months of AI Coding Did to My Dev Team.pdf (60.4 KB)
