Better references beat better prompts
You still need to be crafty with the initial context. Most people skip it. Research papers, niche blog posts, specific docs, reference implementations. Feeding the right material in is the highest-leverage move with AI coding tools.
Based on workflow patterns from building with Claude Code and Codex throughout 2025 and 2026.
The prompt matters less than what surrounds it
Most people spend an hour refining instructions when they should spend fifteen minutes finding the right reference. The best input is not a better prompt. It is a better reference. Give the agent a concrete example of good and the output beats any amount of instruction.
I built ascii.blode.co in a day because I started from a blog article on ASCII rendering. The article was the context. I built allmd to turn any URL, PDF, video, or audio into markdown for an agent. Feed the model a post about the exact technique you want. Watch the quality jump.
Front-load the right material
My Raycast snippet "zcc" expands to: "Do extensive research. Make a plan with phases and todos. Use a swarm of subagents and teams." Even that is context gathering. Tell the agent to research before it builds.
Best results never come from the first attempt. They come from the attempt where I front-loaded the right material.
Curate inputs, not instructions
I use Brian Lovin's /simplify skill at the end of every session. Another context move. It gives the agent a framework to evaluate its own output.
The people who get the most out of these tools are not writing sophisticated prompts. They curate inputs. The gap between mediocre and great is almost never the model. It is the fifteen minutes you did or did not spend gathering context before you hit enter.