Fixed packages
From $3K · free intro
nyk
agent ops · solana · oss
blog · category
Measured implementation, context architecture, and local stacks.
Teams burn 30-40 minutes per Claude session re-explaining what they already knew. A three-tier memory system that makes output compound instead of plateau.
One system asks an LLM to read filings. Another waits for an LLM before it submits an order. Only one puts the model where uncertainty is useful.
The problem was never the window size. It was context rot - performance degrades as conversations fill, not because the model forgets, but because signal drowns in noise.
An arXiv paper documented a Claude Code project with 26,000 lines of context architecture - more instructions than actual code. Their agents stopped hallucinating.
Everyone says 'fix prompts' - in practice, the bottleneck is shell init, hook chains, and context drift. A measured pass cut zsh startup from 1.794s to 0.386s.
newsletter
Free · no spam
Implementation notes on reliable systems, local stacks, and the controls behind repeatable delivery. Hands-on later: packages from $3K.
Engineering notes + checklist. No spam.
