给 AI 交代了一长串要求，结果还是差点意思；收集了不少提示词和技能，真正用起来又总觉得不顺手。到底是模型能力不够，还是我们的用法也需要调整？
最近读到 Anthropic 两位工程师的三篇文章，刚好从三个角度谈到了这个问题。先看这张总结图，再分别展开。
根据参考图重绘，图中篇名为中文概括。前两篇来自 Thariq Shihipar，第三篇来自 Felix Rieseberg，原文链接见文末。
一、提问技巧：找到你的未知
Thariq 提到了一个比喻：交给 AI 的提示词、背景资料，是一张地图；实际需要完成的工作，才是它要走进去的真实世界。地图没有标注的地方，AI 就得根据自己的理解做判断，任务越复杂，这样的地方也越多。找到你的未知原文
所以，需求写得长，不代表我们真的把事情想周全了。比如让 AI 帮忙修一张照片，懂摄影的人可以说出曝光、白平衡、肤色应该怎么调；不懂的人大概率只能说“这张有点灰，帮我调高级一点”。这里的“高级”，可能看到之后就能判断喜不喜欢，但开口之前，确实不知道该怎么描述。更麻烦的是，有些影响观感的因素，我们压根就没注意到。
Thariq 把这些情况分为四类：
- 已知的已知：明确知道照片要发给谁、用在哪里。
- 已知的未知：知道自己不懂调色，需要先了解。
- 未知的已知：心里有偏好，但看到具体版本才想起该怎么描述。
- 未知的未知：连哪些因素会影响效果，都还没意识到。
他建议通过盲点检查、访谈、原型等方式找出遗漏，而且在执行过程中也继续澄清新问题。
放到刚才修图的场景里，可以先让 AI 解释这张图有哪些问题、有哪些调整方向，顺便把不懂的词讲清楚；如果有些偏好实在说不出来，就让它做几个差异明显的版本，看完再选。拿着具体的东西交流，总比反复说“再高级一点”强。
我觉得这个思路最有用的地方，是让人不用等到能写出一份完美的要求，才开始使用 AI。我们找设计师、编辑合作，本来也需要来回沟通，为什么到了 AI 这里，就总期待一句话把它安排明白呢？它问问题、给选项的过程，也可以帮助我们把自己的想法理顺。
不过，共同讨论出来的答案也得检查。修图就看实际导出的成品，查资料就回到原始来源核对。聊得顺畅和事情做对，终究是两回事。
二、上下文工程：给规则做减法
“上下文”可以理解为 AI 干活时拿到的整套材料，除了当前这句话，还包括系统提示词、长期说明、记忆、Skill（可复用的任务指南）等。
Thariq 在第二篇文章里提到，Anthropic 团队为更先进的模型，删掉了 Claude Code 超过 80% 的系统提示词，在他们自己的编码评测中，没有观察到可测量的损失。原因之一是，有些限制是过去模型能力不足时加上的，现在却可能妨碍它发挥；来自不同地方的指令，也可能互相冲突。上下文工程原文
这有点像带新人。刚入职时，把流程交代得细一点很正常。但人家已经能独立负责事情了，还要求他照着入职第一天的小抄干活，能力也就只能表现到那份小抄允许的程度。人可能会有意见，AI 则可能认真地把一堆没必要的步骤走完，再礼貌地告诉你，任务已完成。
尤其有些提示词是从不同地方拼起来的，每一段单独看都很有道理，放到一起却未必合适。一会儿要求尽可能简洁，一会儿又要求任何观点必须充分展开；既要自然口语化，又要每节都有总结。最后的文章句句像结论，读着却累，这也未必全是模型的问题。
原文建议让上下文保持轻量，需要时再加载相关信息。不过，“删掉八成”这个数字不能直接照抄。他们有具体的模型和评测作为前提，而你那些只有自己知道的特殊要求，仍然得交代。文章写给谁看、哪些内容不适合公开，这些事情不会因为模型变聪明，就自动出现在它脑子里。
真要动手清理，可以先找一条过去为了解决某次错误加上的要求，去掉它，用几件有代表性的任务做个对照：原来的问题还会不会出现，有没有少一些无用的步骤，结果是否更好。这样至少知道自己删掉的是什么，不至于把“精简上下文”又变成一次凭感觉的折腾。
三、心智的野心：目标可以再大一点
第三篇把问题往前推了一步：我们想让 AI 做的事情，会不会从一开始就小了？
Felix 把使用 AI 的野心分为速度、高度、深度和广度。速度很好理解，同样的事情做得更快，也是最容易想到的一种用法。其他三个方向，则值得多琢磨一下。更有野心原文
高度，是再往上想一层，自己最终要什么。 比如把文档转成十页 PPT，这是一个明确的任务。但这场演示究竟是为了介绍事情，还是为了让会上做出一个决定？如果是后者，那更值得先做的，也许是从材料里找出需要讨论的分歧、缺失的证据。至于最后到底做几页，可以往后放一放。
深度，是把专业的方法真正用上。 让 AI 总结一本书，和让它帮助自己理解一本书，需要做的工作就不一样。围绕核心判断找反例、换一个立场讨论、出几个必须理解内容才能回答的问题，再把答不上的部分带回原文看，这就比收下一份摘要多走了几步。
广度，是尝试自己不熟悉的领域和工具。 做演示时，除了让它整理文字，也可以尝试用图表解释数据，用动画展示过程，或者做一个能现场操作的小模型。有些想法过去因为自己不会、找人做又麻烦，刚想到就放弃了。现在可以重新评估一下，至少先看看能做成什么样。
上面这些是我顺着文章想到的例子。Felix 还特别提醒，专业的口吻并不等于专业的方法，需要相应的数据、工具和外部验证。
这点很有必要。给 AI 安排一个“二十年经验的专家”身份确实容易，但如果分析行业时没有查过真实数据，点评设计时连实际效果都没看过，挂什么头衔也没用。目标可以放大，解决问题需要的材料和检查手段也得跟上。
最后
这三个观点里，我最认同的是：别太早给自己，也给工具下结论。暂时说不清要求，可以先聊；以前留下来的规矩，可以重新检查；过去觉得做不了的事情，也可以拿出来试试。
做完之后，再看看哪些地方依然没想清楚，哪些限制确实有必要。图里的“完整闭环”，放到实际使用中，大概就是这样不断尝试、修正的过程。模型当然还有很多毛病，但自己的用法是否合适，也值得一起检查。
参考文章：
- Thariq Shihipar：A field guide to Claude Fable 5: Finding your unknowns
- Thariq Shihipar：The new rules of context engineering for Claude 5 generation models
- Felix Rieseberg：How to be more ambitious: A tutorial for humans with models
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
A long list of demands was given to AI, and the result was almost as good; a lot of prompts and skills were collected, and it really didn't work out. Is there a lack of modelling capacity or does our use need to be adjusted?
Three articles by two engineers in Anthropic have recently been read, and the issue has been addressed from three angles. Let's take a look at this summary and move on.
Redrawn from the reference figure, the title of the chart is Chinese. The first two are from Tharik Shihipar and the third from Felix Rieseberg.
I. Questioning technique: finding your unknown
Tharik refers to a metaphor: the prompt given to AI, the background information, is a map; what actually needs to be done is the real world in which it is going to go. Where the map is not marked, AI has to judge according to its own understanding, the more complex the task, the more places. Find your unknown original
So long needs don't mean we really think about things. For example, if AI helps fix a picture, the person who knows the picture can say what to do with exposure, white balance, colour; the person who doesn't know can probably just say, "This one is a little gray, help me make it a little higher." “Advanced”, here, may be able to judge whether or not to like it after seeing it, but it is not clear how to describe it until we speak. Even more problematic is the fact that there are some elements that affect perceptions that we have not noticed at all.
Tharik grouped these situations into four categories:
- Known: know exactly to whom and where to use the photographs.
- Known unknown: know you don't know the color, need to know first.
- Unknownly known: there is a preference in mind, but it is only when you see a specific version that you think of what to describe.
- Unknown: not even the factors that affect the effects.
He recommended that gaps be identified through blind spot checks, interviews, prototypes and that new issues continue to be clarified during implementation.
In the scene of the painting, let AI explain what the problems are, what the adjustments are, and what the intelligence is; if some preferences are not, let it do a few distinct versions and see the selection. It's better to talk with specific things than to say "a little more."
I think the most useful thing about this idea is that people don't have to wait until they write a perfect request to start using AI. We needed to work with designers and editors, and we needed to talk back and forth. The process of asking questions and giving options can also help us to rationalize our thinking.
However, the answers discussed together must also be examined. The chart looks at the actual manufactured product and the information is checked back to the original source. Having a good conversation and doing the right thing is different.
context Project: Subtracting Rules
“context” can be understood as a package of materials obtained during AI's work, including, in addition to the current sentence, the system prompt, the long-term description, memory, Skill (reusable mission guide).
In the second article, Tharik mentioned that the Anthropic team, for a more advanced model, had deleted the prompt system of Claude Code, which exceeded 80 per cent, and had not observed measurable losses in their own coding assessments. One reason for this is that some of the constraints that have been added to the past when modelling capacity was inadequate may now prevent it from functioning; instructions from different locations may also conflict with each other. Original context Project
It's kind of like bringing new people. It is normal to submit the process in a more detailed manner upon entry into service. However, he has been able to take charge of the matter independently and has been asked to work as a small copy of his first day in office, so that his ability can be demonstrated only to the extent permitted by the small copy. People may have an opinion, and AI may seriously finish a bunch of unnecessary steps and then politely tell you that the mission has been completed.
In particular, there are prompts that are co-mingled from different places, and each paragraph alone makes sense and may not be appropriate. A brief call is made for being as concise as possible and for any point of view to be fully developed; it requires both natural verbalization and a summary of each section. The last sentence of the article, which is like a conclusion, is tired of reading, which is not necessarily a question of models.
It was suggested that context should be kept light and that the relevant information should be loaded when needed. However, the figure of 80 per cent could not be copied directly. They have specific models and assessments as a prerequisite, and your special requirements, which only you know, still have to answer. To whom the articles are written and what is inappropriate for disclosure, these things do not automatically appear in the brain of a model because it is smart.
If you really want to clean it up, you can start by looking for a request that was added to one of the previous mistakes, remove it and compare it with a few representative tasks: The original question was whether there would be any less useless steps and whether the result would be better. At the very least, knowing what has been deleted does not make it possible to turn the “simplified context” into a sense-based manoeuvre.
III. Mental ambition: the goal can be bigger
Part three takes the question one step further: Do we want AI to be small from the beginning?
Felix divides the ambition to use AI into speed, height, depth and breadth. The speed is well understood, and the same thing is done more quickly and is one of the easiest uses to think about. The other three directions deserve more thought. More ambitious.
Altitude, one more layer, what it's all about. It's a clear task, for example, to convert the document into a 10-page PPT. But is this presentation about telling things, or is it about making a decision? In the latter case, it would be more worthwhile to do so first, perhaps by identifying from the material the differences and missing evidence to be discussed. As for the final few pages, they could be left behind.
Depth, is to really use professional methods. Let AI sum up a book, and let it help itself understand a book. It would be a few steps more than the next summary to turn around the core judgement, to discuss a different position, to ask a few questions that must be understood in order to answer, and to take the unaccepted part back to the original.
Wideness is an area and tool that is not familiar to you. In doing the presentation, in addition to allowing it to organize the text, it is possible to try to interpret the data with a chart, to show the process with an animation, or to make a small model that can be fielded. In the past, some ideas were given up because they could not be done, because they could not be done. Now we can re-evaluate, at least see what we can do.
These are the examples I came up with following the article. Felix is also particularly reminded that a professional tone is not equivalent to a professional approach and requires corresponding data, tools and external validation.
That is necessary. It is easy for AI to have an “expert of 20 years' experience” as an expert, but if the analysis of the industry has not been accompanied by real data, the evaluation design has not even seen the actual results, and the title is useless. Targets can be amplified and the material and inspection tools needed to solve the problem must be kept up.
Finally.
Of these three points, I agree most: don't be too early for yourself, but for the tools. For the time being, it is possible to talk first; the rules that were left behind can be re-examined; and what was felt to be impossible in the past can be taken out.
After that, we will see where we still do not have a clear picture of what restrictions are really necessary. This is probably the process of continuous attempt and correction. There are, of course, many problems with models, but the appropriateness of their use is worth examining together.
References:
- Thariq Shihipar：A field guide to Claude Fable 5: Finding your unknowns
- Thariq Shihipar：The new rules of context engineering for Claude 5 generation models
- Felix Rieseberg：How to be more ambitious: A tutorial for humans with models
Auto-fetched and translated by AI Atlas. All rights belong to the original author.