Meta-prompting ChatGPT to write one-shot prompts for Work/Astra still takes 2-4 retries
NewPainting8224 · reddit · 2026-09-28
A heavy ChatGPT user shares their struggle getting one-shot prompts right for Work/Astra/Sol: they describe the project to regular ChatGPT, have it write the prompt, but Astra still misses key requirements, forcing 2-4 correction rounds. Iterating, they distilled hard-won prompt elements: preserve what works, don't rebuild from scratch, actually use all provided info, anticipate failure points, define "done", self-audit before stopping, and don't turn redesigns into cosmetic tweaks. Since Work usage limits make trial-and-error costly, they're looking for a reusable meta-prompt that produces the final prompt in one shot — a practical look at translating user intent into high-quality agent instructions.
More from coding & agent
- Claude Code pay-as-you-go credits burn $50 in 15 minutes, user warns against buying extra usage — RileyRalmuto · 2026-09-28
- Don't invert that matrix: the classic numerical linear algebra advice still holds — tdhopper · 2026-09-28
- Open-source eval workbench: test suites plus production monitoring for LLM apps — Vin-builds · 2026-09-28
- Open-source VelaTerm puts Codex-style agent sessions and split panes in one native app — george-lin · 2026-09-28
- I built an AI trading agent, then built one that doesn't use AI to trade — murdocc007 · 2026-09-28
- Anthropic researcher had Claude build a custom visualization of their giant codebase — gallabytes · 2026-09-28