用 Jev 先筛掉坏方案，再让 Opus 5.5 实现以省 90% 成本

Gresham @Greshamtrade原推发布于 2026.10.16

在 X 上打开原推

THIS IS STUPIDLY F*CKING GOOD Anthropic engineers have a private playbook with a rule for cutting Opus 5.5 costs by 90% that almost nobody outside the company knows. The idea: make Opus propose 5 ways to solve a task, then let Jev eliminate the weak ones before you pay Opus to implement anything. Most agents pick their first idea, start building, hit a problem, backtrack, and try again. You pay for every wrong turn. This workflow rejects bad approaches while they’re still a few lines of text. You run it like this: task → 5 short plans → Jev checks each one → Opus builds the winner here’s the split: → Opus 5.5 proposes the approaches and names their assumptions → Jev checks each plan against your requirements and project evidence → Code rejects plans that fail a mandatory check → Opus implements the strongest surviving plan You only look at three moments: → when the requirements are set → when no plan survives → before accepting the finished result Opus proposes. Jev challenges. Opus executes. The trick is what you ask Jev: “Does this plan require a dependency we don’t have?” or “Does it change a file we must leave untouched?” Each answer can kill a dead end before you pay to build it. the full tree: > Opus writes 5 short plans with assumptions and a verification step > Jev checks each assumption against the same project evidence > each check returns supported, contradicted, or unknown > code removes plans with a contradicted mandatory assumption > important unknowns trigger a small check before implementation > surviving plans are ranked by scope, effort, and reversibility > Opus builds the winner and verifies the result a failed check reopens the specific assumption behind it Five possible approaches. One expensive implementation. You’re comparing outlines instead of discovering four mistakes in finished code. paste this into your coding agent ↓ “Set up a plan tournament for this task: 1. Use Opus 5.5 to propose 5 distinct approaches, maximum 80 words each. Don’t implement them yet. 2. List each plan’s dependencies, assumptions, affected files, and first verification step. 3. Give Jev the project evidence and check each assumption separately: supported, contradicted, or unknown. 4. Reject plans that contradict mandatory requirements. Resolve important unknowns with the smallest available check. 5. Select the surviving plan with the smallest scope that satisfies the requirements. 6. Use Opus to implement that plan and verify the result. Show me the comparison before implementation.” The complete Opus 5.5 + Jev workflow is in the article below ↓

这条是引用着另一条写的：x.com/0xnineinch/status/2108604801411879139