8 suggestions for structuring code and prompts around the actual architectural limits of Transformer-based AI assistants.
Also: revisiting well established software engineering principles and tools ⚙️💡
Transformers have no call stack, no persistent memory and no dependency tracker. They approximate all of these through attention, and that approximation breaks down past certain depths and distances.
Each suggestion comes with:
- A concrete before/after example from coding workflows
- The specific failure mode it prevents
- The relevant AI research citation
These practical suggestions to work around their limits should apply to any Transformer-based assistant - Claude, GPT, Copilot, or local open-source models like DeepSeek, Qwen, Llama, and Mistral.
They should help more on local models because the limitations hit sooner.