When Trying to Make AI Better Makes It Worse
A million words of failed drafts taught me that improving my AI setup sometimes means giving it fewer rules to follow
TL;DR
- The author experienced 'Verschlimmbesserung' by adding too much context and rules to AI prompts, leading to degraded performance.
- Attempts to 'save everything', 'connect everything', and 'codify everything' made the AI's task of generating content overwhelmingly complex.
- The excessive context included 3,855-word style guides and running logs of every correction, causing the AI to follow rigid, often contradictory, instructions.
- To fix the issue, the author reset the AI context, discarding the complex tangle and rebuilding with significantly shorter, principle-based guides.
- Key principles for successful AI context include the agent asking first, each folder standing alone, and keeping guides concise.
- Rebuilding trust in the AI and oneself is crucial, involving careful context management and avoiding major decisions under pressure.