AI PROMPT ·
在 ChatGPT/Codex 中使用 Astra 的 15 条规则
"提供排名对比。显示每组的样本量。标记基于极少帖子得出的结论。" 不要说"think step by step"。直接给出答案。 8). 附加文件。解释其作用。 ↳ 确切提示词："Use the brand guide for style and the product brief for facts. If they disagree, flag the conflict first." 证据、模板或指令：Astra 无法区分。 9). 说明哪些决策需要你的输入。 ↳ Astra 比任何之前的模型都更频繁地请求澄清。这对头脑风暴很有用，但对于你已有解决方案的任务来说非常烦人。列出值得询问的 2-3 个问题。 10). 让它处理其余部分。 ↳ 确切提示词："Otherwise, complete the plan and list any assumptions at the end." 11). 指定语气、长度和格式。 ↳ 默认情况下，句子过于完美。"600-800 words. Paragraphs of 2-3 sentences. Answer Greg's question in the first line." 我最喜欢的技巧："Use ASD-STE100." 12). 添加："List your assumptions." ↳ 当它试图填补空白时，你可以很容易地发现它的猜测。尽早发现错误信息。 13). 在它工作时纠正方向。 ↳ 它在长任务上表现相当出色，同时消耗 token。因此，在任务中途发送后续消息以查看其计划和执行情况。就像……对待人类一样？ 14). 明确告诉它要检查什么。 ↳ 确切提示词："Before delivering, confirm the totals match the source, identify duplicates, and check every chart uses the correct date range." 15). 添加："Flag anything unresolved." ↳ 这消除了虚假的 "all done!" 报告。未验证的内容保持未验证状态。
PROMPT
"Deliver a ranked comparison. Show the sample size for each group. Flag conclusions based on very few posts." No "think step by step". Be direct instead. 8). Attach files. Explain their role. ↳ Exact prompt: "Use the brand guide for style and the product brief for facts. If they disagree, flag the conflict first." Evidence, template, or instructions: Astra can't tell. 9). Say which decisions need your input. ↳ Astra asks for clarification more than any model before. Useful to brainstorm but it gets very annoying for the tasks you already had a solution for. Name the 2-3 questions worth asking. 10). Let it handle the rest. ↳ Exact prompt: "Otherwise, complete the plan and list any assumptions at the end." 11). Specify tone, length & format. ↳ The default is overly perfect sentences. "600-800 words. Paragraphs of 2-3 sentences. Answer Greg's question in the first line." My favorite trick: "Use ASD-STE100." 12). Add: "List your assumptions." ↳ You can catch its guesses very easily when it tries to fill the gap. Catch the misinformation early. 13). Correct its direction as it works. ↳ It is quite good at long tasks, while consuming tokens. So send a follow-up mid-task to see its plan & execution. Just like... a human? 14). Tell it exactly what to check. ↳ Exact prompt: "Before delivering, confirm the totals match the source, identify duplicates, and check every chart uses the correct date range." 15). Add: "Flag anything unresolved." ↳ This removes the fake "all done!" report. Unverified stays unverified.