Built an AI system to manage my day. Noticed the AI drops balls the same way I do: forgets instructions from earlier in the conversation, rushes to output, skips boring steps.
Research confirms it:
- "Lost in the Middle" (Stanford 2023): 30%+ performance drop for mid-context instructions
- 65% of enterprise AI failures in 2025 attributed to context drift
So I built scaffolding for both sides:
For the human: friction-ordered tasks, pre-written actions, loop tracking with escalation.
For the AI: verification gate that blocks output if required sections missing, step-loader that re-injects instructions before execution, rules preventing self-authorized step skipping.
Open sourced: https://github.com/assafkip/kipi-system
README has a section on "The AI needs scaffolding too" with the full
research basis.
[link] [comments]