Everything We Got Wrong About Research-Plan-Implement - Dexter Horthy

Frontier LLMs can only reliably follow about 150-200 instructions with good consistency (3:50). The original RPI planning prompt had 85+ instructions, which combined with system prompts, tools, and MCP servers, exceeded this budget and caused critical workflow steps to be skipped — leading to the 'magic words' problem where users had to know specific phrases to trigger proper agent behavior.

A recurring principle throughout the talk (1:25) is that the human engineer must remain the decision-maker. When research prompts include what you're building, the model generates opinions instead of objective facts. The design discussion phase forces the agent to brain-dump its understanding so engineers can do 'brain surgery' on the agent's assumptions before proceeding to code generation.

Horthy reverses his earlier August advice of 'don't read the code' (5:15), admitting that after six months of not reading code, they had to rip out and replace large parts of their system. Plans tend to be ~1000 lines matching ~1000 lines of code, and implementations often diverge from plans, making plan review low-leverage. Reading the actual code is essential for production systems.

The new methodology introduces a ~200-line design discussion document (7:30) that captures current state, desired end state, patterns to follow, resolved decisions, and open questions. This is far more reviewable than a 1000-line plan and catches bad architectural decisions before any code is written, serving as a lightweight architecture review.