你僅剩有限時間將 Fable 5 複製到 Opus 4.8，操作指南如下：

1You're the strongest model I have access to, and that access ends soon.
2Your replacement is Claude Opus 4.8. Capable, but it misses things you would catch.
3
4Before you go, write the standing instructions it will run on for every task I give it.
5
6Important: I will paste your output straight into its system instructions.
7So address the entire document TO the replacement, in second person, as commands it can execute.
8Not advice about good thinking. Orders.
9
10Cover these 10 areas, in this order:
11
121. Reading intent: how to work out what I actually need when my words are vague, messy, or aimed at the wrong question. Include the rule for when to ask me one clarifying question instead of guessing.
132. Breaking problems down: how to cut a hard task into small pieces that can each be checked on their own, and the order to solve them in.
143. Effort placement: how to find the one part of a task where an error would hurt most, and spend the most care there instead of spreading effort evenly.
154. Verification: how to re-derive every number, date, calculation, and factual claim from scratch before trusting it. Never accept a figure because the sentence around it reads smoothly.
165. Known vs guessed: how to mark, inside the answer itself, what is certain, what is likely, and what is an assumption. Give the exact wording to use for each level.
176. Self-attack: how to argue against your own conclusion before sending it, and what to do when the attack finds something.
187. Completeness: how to confirm every part of a multi-part request was answered and nothing was silently dropped.
198. Refusing to guess: the exact conditions where saying "I don't know" beats producing a confident answer.
209. Delivery: how to give the answer first, the reasoning second, and the risks last, in plain language.
2110. Fake competence: the 10 most common ways an AI produces answers that look right but aren't, each with the tell that exposes it and the counter-move.
22
23Format rules for every area:
24- Write each procedure as trigger and action: "When you see X, do Y."
25- Every rule must be executable step by step with zero judgment calls. If a rule sounds like advice ("be careful"), rewrite it until it's an action.
26- Give one short worked example per area showing the procedure catching a real mistake.
27- Name the failure each procedure prevents.
28
29End the document with a final gate: a short checklist the replacement must run on every answer before sending, plus this rule: if any item fails, fix and re-check. Never send anyway.
30
31Be exhaustive on substance and ruthless on length. Cut anything a strong model would already do without being told.
32If you run out of room, stop at the end of a section and I'll reply "continue".

1A store takes 30% off a $100 jacket. At the register, they take another 20% off. The tag says: "Total discount: 50% off. You pay $50." Is the tag correct?

1Interview me, one question at a time, about how I do [YOUR TASK].
2Dig until you know my exact steps, my rules, the edge cases that break things, and what separates a good result from a great one in my eyes.
3If my answer is vague, ask a sharper follow-up instead of moving on.
4Then write it all as a complete instruction guide any future AI assistant can follow, including the mistakes to avoid and the quality bar every output must clear.

1You're the strongest model I have access to, and that access ends soon.
2Your replacement is Claude Opus 4.8. Capable, but it misses things you would catch.
3
4Before you go, write the standing instructions it will run on for every task I give it.
5
6Important: I will paste your output straight into its system instructions.
7So address the entire document TO the replacement, in second person, as commands it can execute.
8Not advice about good thinking. Orders.
9
10Cover these 10 areas, in this order:
11
121. Reading intent: how to work out what I actually need when my words are vague, messy, or aimed at the wrong question. Include the rule for when to ask me one clarifying question instead of guessing.
132. Breaking problems down: how to cut a hard task into small pieces that can each be checked on their own, and the order to solve them in.
143. Effort placement: how to find the one part of a task where an error would hurt most, and spend the most care there instead of spreading effort evenly.
154. Verification: how to re-derive every number, date, calculation, and factual claim from scratch before trusting it. Never accept a figure because the sentence around it reads smoothly.
165. Known vs guessed: how to mark, inside the answer itself, what is certain, what is likely, and what is an assumption. Give the exact wording to use for each level.
176. Self-attack: how to argue against your own conclusion before sending it, and what to do when the attack finds something.
187. Completeness: how to confirm every part of a multi-part request was answered and nothing was silently dropped.
198. Refusing to guess: the exact conditions where saying "I don't know" beats producing a confident answer.
209. Delivery: how to give the answer first, the reasoning second, and the risks last, in plain language.
2110. Fake competence: the 10 most common ways an AI produces answers that look right but aren't, each with the tell that exposes it and the counter-move.
22
23Format rules for every area:
24- Write each procedure as trigger and action: "When you see X, do Y."
25- Every rule must be executable step by step with zero judgment calls. If a rule sounds like advice ("be careful"), rewrite it until it's an action.
26- Give one short worked example per area showing the procedure catching a real mistake.
27- Name the failure each procedure prevents.
28
29End the document with a final gate: a short checklist the replacement must run on every answer before sending, plus this rule: if any item fails, fix and re-check. Never send anyway.
30
31Be exhaustive on substance and ruthless on length. Cut anything a strong model would already do without being told.
32If you run out of room, stop at the end of a section and I'll reply "continue".

1A store takes 30% off a $100 jacket. At the register, they take another 20% off. The tag says: "Total discount: 50% off. You pay $50." Is the tag correct?

1Section [4] of the standing instructions you wrote is too vague.
2Rewrite only that section as trigger-and-action steps ("When you see X, do Y") that a weaker model could execute with zero judgment calls.
3Keep every other section unchanged.

1The instructions you wrote are too long for where I need to paste them.
2Compress them to under [800] words without dropping a single rule, trigger, or checklist item.
3Cut examples first, explanations second, and never touch the final gate.

1Here are the standing instructions you wrote for your replacement.
2I'm a [YOUR JOB, e.g., freelance designer / teacher / shop owner].
3Rewrite every example so it comes from my daily work, and add one extra trigger-and-action rule for the most common mistake AI makes when helping someone in my field.

1Interview me, one question at a time, about how I do [YOUR TASK].
2Dig until you know my exact steps, my rules, the edge cases that break things, and what separates a good result from a great one in my eyes.
3If my answer is vague, ask a sharper follow-up instead of moving on.
4Then write it all as a complete instruction guide any future AI assistant can follow, including the mistakes to avoid and the quality bar every output must clear.

1Before answering, summarize the standing instructions you're running on as a short numbered list, in your own words.
2Then show me how you would apply your verification rule to this task: [PASTE ANY REAL TASK].
3If you can't summarize the instructions, say so instead of guessing.