跳到正文
DEV Community· Devesh Korde·· 8 天前AI 評分55

Claude 四種提示詞模式值得測試,作者不作 10 倍效果宣稱

Four Prompt Patterns Worth Testing on Claude (None of Them Is 10x)

AI 導讀

Devesh Korde 整理四種可在 Claude 上測試的提示詞模式,並明言沒有 10 倍提升的依據。其中三種有 Anthropic 公佈的測試結果:範圍控制提示在 max effort 編碼測試中令工作階段成本下降約三分之一且品質不變;要求實際檢查則令 low effort 下略過或流於表面的檢查變得罕見,但每項任務成本略升。

正文

My working title for this post was "The Prompt Pattern That Gets 10x Better Results from Claude."

I changed it. I can't back that up.

A 10x claim needs a metric. Ten times fewer errors? Ten times cheaper? Without a metric, a multiplier is a slogan, and nobody can check it.

What I can show are four patterns Anthropic documents, three of them with published test results.

Scope: I did not run any benchmarks. The results below are Anthropic-reported, mostly measured on Claude Sonnet 5.5, from its prompting guides, read on 2026-11-10. The prompt wording in the full post is mine and untested.

Four boxes labeled task, scope, proof, shape

The idea: a prompt is a contract

Most prompts say what to do and stop. A better one makes four promises explicit: task (what and why), scope (where the job ends), proof (how the model shows it's done), and shape (what the answer looks like).

1. Scope control

Per Anthropic's Sonnet 5.5 guide, at low and medium effort on long agentic tasks the model sometimes checks in before finishing. At every effort level it tends to add tests, docs, and small files you didn't ask for.

At xhigh and max effort it can also start its own review rounds, sometimes with reviewer subagents. Anthropic reports that an instruction to stop once the work and its checks pass, with no extra review rounds unless requested, cut session cost by about a third in coding tests at max effort, with no change in quality.

Read the fine print: one model, one effort level, one family of tasks, one specific prompt. Anthropic also notes the instruction reduces self-started reviews without removing them entirely.

A runner stopping at the finish line instead of running past it

2. Proof of work

At low effort, Sonnet 5.5 sometimes reports a code change as done without running a check that exercises it, for example skipping the tests because dependencies weren't installed.

Anthropic reports that its version of a "run a real check" paragraph makes skipped or superficial checks rare at low effort, with no measurable quality change and a slightly higher cost per task. The pattern has three parts: require a real check, define what counts as real, and give an honest exit ("say which check you couldn't run and why").

3. The reasoning nudge

This one is narrow. It applies to JSON answers on tasks that need a few steps of working out, like totaling figures or applying a rule. With structured outputs, the model can only work the problem out in its thinking, and it often skips thinking at low and medium effort.

Anthropic's fix is one line at the end of the system prompt, used with adaptive thinking:

Think the problem through before you answer.

Reported results: at high effort it brings accuracy close to xhigh for a modest token increase. At low and medium it raises accuracy but not to high's level, with a larger token increase. It does nothing with between_tools, because that setting means no up-front thinking. And treat stop_reason: "max_tokens" as a failed response, even if the JSON parses.

A dial turned from low to high next to a thought bubble

4. The output contract

No published numbers for this one. It's guidance: name the format positively, use tags for the parts you parse, and don't ask for the model's reasoning in the answer. On several current models, including Sonnet 5.5, that can trigger a reasoning_extraction refusal. Read summarized thinking blocks instead.

What I'm not claiming

  • I didn't reproduce any of these results. They're Anthropic's, on specific models and tasks.
  • They don't transfer automatically. I don't know how Haiku 5.5 or Opus 5.5 behave.
  • Only three of four patterns have numbers. The output contract is guidance.
  • There is no 10x. If someone shows you one, ask what they measured.

Test it on your own inputs

The full post has what this teaser doesn't:

  • Complete, copyable system prompt blocks for scope, proof, and output format
  • A small Python harness that compares prompt variants on pass rate, output tokens, and latency
  • The "prompts decay" section: which old instructions to delete when you change models

Read the full post: Four Prompt Patterns Worth Testing on Claude

This is Day 6 of a series on Claude foundations. Day 5 covered XML tags for structuring prompts.

來源:DEV Community · dev.to