A Two-Stage Trick for GPT-6 Astra: Build Demos on Medium, Polish on Max

omarsar0 · x · 2026-09-06

omarsar0 shares an iterative workflow that works well on GPT-6 Astra: feed it a tweet of an impressive Astra demo, have the medium tier replicate it as best it can with extra instructions, and set a /goal (e.g. "provide proof of results") so it has something to compare against; then switch to the max tier to polish, repeating as needed.

He argues it works because it decomposes the problem — one component builds, the other optimizes — instead of burning tokens on many things at once, which usually yields lower quality. The whole flow can be done in one pass with subagents and /goal.

Original post →

More from coding & agent

coding & agent channel →