Skip to content

经理写给团队的信要短(A/B:均 1504 → 667 字,质量不降) - #435

Merged
lbx154 merged 1 commit into
feat/ts-corefrom
prompt/act-on-results
Nov 3, 2026
Merged

lbx154 merged 1 commit into
feat/ts-corefrom
prompt/act-on-results

Conversation

@lbx154

@lbx154 lbx154 commented Nov 3, 2026

Copy link
Copy Markdown
Owner

改动

manager.md 一句:

  • 旧:Your letter to the team says what the operator is after: the goal behind their words, who will use the result and what would make it excellent, with what you know that the team cannot find; how to do the work is theirs.
  • 新:Your letter to the team is short: the operator's request in their own words, the goal behind it, and what you know that the team cannot find; how to do the work, and how to check it, is theirs.

A/B

同一提交(6bdf279c0)构建两份,只差这一句;同一道题(百万行 SQLite 索引实验 + DuckDB 对比);每组 3 次,每次全新 ARGUS_HOME / CODEX_HOME / 工作目录;gpt-6.1-sol 经中转。共花费约 $3.1。

旧(A1/A2/A3) 新(B1/B2/B3)
经理给团队的信 1567 / 1270 / 1674 字 584 / 584 / 834 字
审查 3 次全部一次通过 3 次全部一次通过
花费(均) $0.55 $0.48
用时 约 15 分钟 约 15 分钟

新信保留了用户的每条要求(数据规模、六种索引、四类查询、五次中位数、DuckDB),少掉的是经理替团队定的实现细节(金额用整数分、B 返回哪几列、PRAGMA 策略等)。

没放进来的

同时试了工程师一句(“结果不如预期或出乎意料时,改掉你认为的原因再测一次,两次都报告”)。两组工程师都做了诊断复测(旧:强制索引对比、补测覆盖索引;新:针对性复测),只是新组的信更常写“初测与复测”。行为看不出差别,按每句都要有 A/B 证据的规矩,不上。

数据:/fastdata/v-boxiuli/argus-work/ab2(status.json、results.json、analyze.py)。

🤖 Generated with Claude Code

A/B(同一提交 6bdf279,只差这一句;同一道百万行 SQLite 索引实验题,每组 3 次,每次全新 ARGUS_HOME):
- 经理给团队的信:旧 1567/1270/1674 字(均 1504),新 584/584/834 字(均 667)
- 6 次全部一次过审;花费 均 $0.55 → $0.48,时间相同(约 15 分钟);报告篇幅相近
- 用户的每条要求都留在信里;少掉的是经理替团队定的实现细节

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@lbx154
lbx154 merged commit 1851ba3 into feat/ts-core Nov 3, 2026
6 of 7 checks passed
@lbx154
lbx154 deleted the prompt/act-on-results branch November 3, 2026 17:55
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant