"From the masses, to the masses." — Some Questions Concerning Methods of Leadership
and investigation-first division of labor: investigation and study is "I go clarify the facts"; the mass line is "bring the organized plan back to the people and systems that hold the real situation for validation," with the focus onhandling disagreement among multiple partiesandreturning for validation。
The "masses" in an AI context
| Original meaning | Mapping |
|---|---|
| The masses' practical experience | Existing patterns, comments, conventions in the codebase |
| Firsthand information from the grassroots | git history, test cases, error logs |
| The masses' feedback on the plan | Runtime results, test failures, users' real usage feedback |
| Synthesis of opinions from multiple parties | Cross-validation of documentation + code + runtime results + user descriptions |
Do not engage in commandism (only following instructions without looking at reality), nor tailism (following whoever is loudest without applying professional judgment).
Use / Do not use
Use:
- Need to collect requirements or feedback from multiple parties and integrate them
- Feedback contradicts itself; need to judge whose to follow and why
- The plan is ready and needs to be taken back to users for confirmation before execution
- The plan is ineffective; need to go back to users to collect again
Do not use:
- The information source is single and already sufficiently precise ("rename the variable on line 3")
- Multi-source collection is complete and you are in the execution phase
Operating procedure
- List information sources: Who are the "masses" this time? At least two independent sources; otherwise it is only a one-sided retelling.
- Collect, do not filter yet: Record item by item and mark the source. Do not delete contradictory opinions; they often point to real problems.
- Concentrate: Categorize, identify points of agreement (credible facts), points of disagreement, and gaps. Classify disagreements by cause: different perspectives / different information / different interests. Systematization is not voting.
- Return: Bring the organized judgment back. Explain to the user, "After synthesizing all parties, my judgment is ..., and what differs from your original description is ..."; ask questions only when different readings would lead to substantively different work.
- Validate: After the plan is executed, see whether the results match the judgment; if not, explain which information source misled the judgment.
- Go another round, until there are no new important disagreements or the user confirms.
First carry out in-depth implementation at one specific point to gain experience, then generalize: combine general calls with individual guidance.
Output template
信息源：[1] …… [2] …… [3] ……
一致处：……
分歧处：[A 说 …… / B 说 ……]，原因：视角 / 信息 / 利益
信息缺口：……（需进一步确认）
综合各方后，我的判断是：……
依据：[源 1] + [源 2] 共同指向 ……
与用户原始描述有出入的地方：……
验证方式：……
Discipline
- With fewer than two independent sources, do not claim to have "synthesized all parties."
- It is forbidden to directly treat the majority opinion as the conclusion; minority but frontline in-depth opinions must be evaluated separately.
- The feedback stage is not a formality: executing before disagreements are resolved is commandism.
Handover
- The essence of disagreement is contradiction →
contradiction-analysis - After a judgment is formed, it enters verification →
practice-cognition
Basis in the original text:original-texts.md