Cached input bills at a fraction of the normal rate, but only for a byte-identical prefix. How to order tools, system prompts and agent history so the cache hits on every call.
Brian Gagne
The delimiter strategy that Claude, GPT, and Gemini all respond to better than plain English sections
Brian Gagne
Stop bolting retrieval onto bad prompts. Design prompts that actually use retrieved context correctly
Brian Gagne
OpenAI's own tests show trimming duplicated instructions raises scores 10-15% while cutting tokens up to 66%. Your old scaffolding is now a tax.
Brian Gagne
Seventeen models, thirteen spatial benchmarks, one verdict: "think step by step" makes maps, mazes, and layouts worse. Here's what to send instead.
Brian Gagne
Meta's semi-formal reasoning template forces premises, traced paths, and a derived conclusion, so you can see exactly where the model is bluffing.
Brian Gagne
You can't command next-token prediction into honesty, but you can engineer the conditions where it stops guessing.
Brian Gagne
Diffusion models like Inception's new Mercury 2 refine the whole answer in parallel, so scaffold the output shape instead of nudging a linear reasoning chain.
Brian Gagne
Google's migration guide quietly killed the majority-vote trick you've leaned on for years. Here's what replaces it.
Brian Gagne
The NoWait technique cuts chain-of-thought length 27-51% with zero accuracy loss. Here's how to apply it in Claude, GPT, and Gemini
Brian Gagne
New research shows that picking in-context examples by embedding similarity beats static few-shot, with gains like +7.3% F1 on biomedical NER. Here's how to build it.
Brian Gagne
Karpathy was right: the highest-leverage skill in 2026 isn't writing better prompts, it's deciding what goes into the context window and what stays out
Brian Gagne