A workflow that makes Astra's "done" verifiable
The author argues against adding more prompts and instead proposes state, decisions and evidence files to turn Astra's "almost done" into a checkable definition of completion.
this is free f*cking gold. how to stop GPT-6 Astra from spending hours on a task and leaving you with nothing you can verify. “working.” “running tools.” “almost done.” none of those tell you what actually changed. the missing piece isn’t another prompt. it’s a workflow that separates three things: • STATE - what’s finished, what’s blocked and the next concrete action • DECISIONS - what was chosen, why, and when it should be reconsidered • EVIDENCE - what test ran, what it returned and what still fails a saved plan isn’t progress. a tool call isn’t success. a file existing doesn’t mean it does the right thing. for a checkout, the payment working is only the beginning. send the callback twice. attempt a refund. simulate a timeout after the charge. then compare the results against the original requirements not the agent’s summary. these files aren’t magic memory. your workflow has to read them, update them and check the evidence. that’s the architecture I break down in the article: persistent state, activity tracking and a finish line defined before the work starts. don’t ask Astra to sound more confident. make “done” harder to claim. read the breakdown before you pay for another hour of “almost finished” ↓
This post quotes another post: x.com/mr_kozh/status/2097221341107925350