클로드 스킬에 포함할 4가지 문장
저는 직장에서 새로운 일을 맡게 되었습니다: 주어진 비구조화된 텍스트 데이터 덤프를 받아 이번 분기 고객이 우리 제품에 대해 말하고 있는지에 대한 상세 PDF 보고서를 제공해 달라는 것이었습니다.
그래서 저는 명확한 프롬프트를 만들었습니다. Claude에게 구체적인 지침을 주었고, 해당 데이터를 입력했습니다. Claude가 결과를 내어 제출했습니다.
하지만 이해관계자와 저는 결과물을 깊이 있게 검토하면서 일부 불편함을 감지했습니다.
클로드는 확신에 차게 틀렸습니다.
Not wrong wrong, like hallucinating facts from nowhere. More like… overconfident wrong. It would generate a quarterly insight report and say something like:
“Negative sentiment in the Dresses department increased 23% this quarter, indicating a significant shift in customer satisfaction that warrants immediate attention from the product team.”
Sounds great. Except that spike was driven almost entirely by a single popular item that launched mid- quarter with a known sizing defect. One product. Not the whole department.
Claude had no idea. And my prompt didn’ t tell it to care.
Image generated by author using Claude
분기별 고객 리뷰 보고서 기술
I’ m going to walk you through a Claude skill I built that generates a quarterly customer sentiment report from unstructured product review text, delivered as a PDF to stakeholders.
Obviously, I’ won’ t be sharing the actual dataset I analyzed at work. The dataset I’m using is the [Women’ s E-Commerce Clothing Reviews ](http://kaggle.com/ datasets/ nicapotato/ womens- ecommerce- clothing- reviews)dataset from Kaggle (CC0 license). It contains 23,000 real, anonymized customer reviews across clothing departments (Tops, Dresses, Bottoms, Jackets, and more) with text, star ratings, and product metadata. References to the company in the reviews have been replaced with “retailer.”
The skill should:
-
Read a filtered slice of reviews for the current quarter
-
Group them by department
-
Identify trends & concerns
-
Write a professional summary PDF for the product leadership team
Here’ s the original prompt:
You are a data analyst generating a quarterly customer sentiment report for a women’ s clothing e- commerce retailer. Given this quarter’ s customer reviews (including review text, star ratings, and department), write a professional stakeholder report that includes:
– An overall sentiment summary for the quarter
– Key themes by department (Tops, Dresses, Bottoms, Jackets)
– 2-3 standout insights from the review text
– A brief recommendation for the product team
Be professional and clear.
When you’ re done with this task, please create a skill titled reviews- analysis and save your instructions in there*.
“신뢰감 있게 틀렸다”는 실제 모습
Here’ s an example of what Claude produced with the naive skill above, on a quarter where the Dresses department had an influx of negative reviews:
“Negative sentiment in the Dresses department increased significantly this quarter, with customers frequently citing fit and sizing issues. This suggests the retailer’ s sizing standards may be drifting from customer expectations — a trend that, if unaddressed, could erode brand loyalty in this key category.”
The real explanation? One dress (a single SKU) launched in Week 7 with a batch quality issue. The reviews were almost entirely about that one item. The rest of the Dresses department was performing fine.
Claude didn’ t necessarily invent anything. It just had no context for why the pattern existed. And without that context, it did what LLMs do: it filled the gap with the most plausible- sounding narrative.
Image generated by author using Claude
Line 1: 클로드에게 필요한 맥락을 알려라
You do NOT have access to product launch calendars, inventory records, promotional campaigns, or individual SKU- level history. Do NOT attribute department- level trends to brand-wide causes. Report patterns you observe in the text; do not explain why they exist unless the reviews themselves make it unambiguous.
This single instruction eliminates a huge category of confident wrongness. Without it, Claude will always reach for a strategic narrative because that’ s what a good analyst does, and Claude is trying to be a good analyst.
The problem is that a good analyst also knows what they don’ t know. They say “We’ re seeing elevated sizing complaints in Dresses this quarter. This may be isolated to a recent launch but we’ d need SKU- level data to confirm.” Claude won’ t say that unless you tell it to.
Line 2: “실질적”이란 무엇을 의미하는지 정의하라
Claude loves the word significant. It uses it all the time. And it almost never defines it.
Only flag a sentiment shift as “significant” if it represents a change of more than 15 percentage points in positive/negative ratio compared to the prior quarter, OR if a theme appears in more than 20% of reviews in a given department. For smaller signals, use language like “slight uptick” or “minor increase.” Do not use the word “notable” or “significant” for anything below these thresholds. Always report the actual number value for the shift along with your claim.
You can adjust the 15% and 20% thresholds to whatever makes sense for your data. The point is to anchor Claude’ s language to something real.
Without this, Claude will call both a 3- review spike in complaints and a genuine 30-point sentiment drop “significant”. Your stakeholders will start to tune out. And when something actually significant happens, they won’ t know it.
Line 3: 모든 통찰에 신뢰 라벨을 부여하라
Before each insight, include a confidence label in brackets: [Data-Supported], [Possible], or [Speculative].
Use [Data-Supported] only when the insight follows directly from the review text provided. Use [Possible] when the insight is a reasonable inference from the text. Use [Speculative] when you are making assumptions about causes or context that are not present in the reviews themselves.
When I first added this line, I was expecting mostly [Data-Supported] tags. What I actually got was a mix of all three, which told me exactly how much Claude had been filling in gaps in my previous reports without me realizing it.
An example of what the output looks like after adding this line:
Image generated by author using Claude
Now your stakeholders can see exactly what’ s solid and what’ s a guess. That’ s a much more honest report.
클로드를 활용해 기술을 개선하는 방법
Writing a skill once isn’ t enough. You need to test it and improve it the same way you’ d iterate on a model.
Step 1: Run the skill on known examples.
Filter the dataset to a time window where you already know what happened. (A quarter with a product recall, a seasonal promotion, a period with unusually high