Claude V3 vs Local IG/RAG V2 对比

以下为模型输出的原始 Markdown，未做内容改写。用于审阅规则本身是否成立，以及后续提炼共同规则。

# Claude V3 vs Local IG/RAG V2 对比

## Compared Drafts

- Claude V3:
  `drafts/008_GOOD_COMPANY_BAD_BUY_CLAUDE_100K_V3.md`
- Local IG/RAG V2:
  `drafts/010_GOOD_COMPANY_BAD_BUY_LOCAL_IG_RAG_RULE_STACK_V2.md`
- Failed local V1:
  `drafts/009_GOOD_COMPANY_BAD_BUY_LOCAL_IG_RAG_RULE_STACK_V1.md`

## V1 Failure

Local V1 failed because it invented a visible frame: `三张账`.

That frame sounded like a model trying to name a structure, not like a writer
trying to explain a real problem. It violated the merged rule stack's own
principle:

- do not replace thinking with a coined framework;
- do not make the structure visible when the reader needs a story;
- do not use abstract labels to simulate clarity.

V1 is kept only as a failed sample.

## V2 Fix

Local V2 removes the artificial frame and returns to the actual reader problem:

> 好公司最容易让人放下戒心。

The article now moves through a more natural sequence:

1. good company creates trust;
2. trust makes price discipline easier to skip;
3. price may have already prepaid future excellence;
4. story feedback can hide operating constraints;
5. the final issue is whether there is enough room for error.

## First-Screen Difference

Claude V3 still has the stronger emotional hook.

- Claude starts with a vivid bad-outcome scene.
- Local V2 starts with a cleaner analytical claim.

Local V2 is less irritating than V1, but still less punchy than Claude.

## Structure Difference

Claude feels more like a finished公众号 essay.

Local V2 is more disciplined and cleaner on the logic:

- it separates company quality from price expectations;
- it adds management-language and earnings-call cues;
- it ends with a six-question checklist.

## Compliance And AI-Flavor Scan

Local V2 scan result:

- non-whitespace characters after metadata stripping: `3004`;
- hard AI-flavor shell hits: `0`;
- direct trading-action hits outside disclaimer: `0`.

Claude V3 is more readable, but had residual trading-flavored wording such as
`买进去/买进`.

## Current Verdict

Claude V3 is still better as a public first read.

Local V2 is now a valid comparison sample. It is cleaner than V1 and much less
artificial, but it still needs a sharper first-screen scene if the goal is to
match Claude's reader pull.

Best next move:

- read Claude V3 and Local V2 side by side;
- decide whether the final system should preserve Claude's opening style while
  keeping Local V2's cleaner compliance and operating-detail discipline.