# Workflow Playbook

## Table of Contents

- 1. Choose the mode and entry point
- 2. Stage 1 checklist
- 3. Clustering rules
- 4. Logic mapping rules
- 5. Stage 2 checklist
- 6. Stage 3 checklist
- 7. One-shot mode
- 8. Common failure patterns

## 1. Choose the Mode and Entry Point

Start from Stage 1 by default.

Choose one of these modes first:

- source-driven: the user provides transcripts, notes, interviews, speeches, or scattered material
- draft-deepening: the user provides a rough draft, partial article, or stable outline that needs depth
- topic-driven: the user provides a topic, thesis, question, angle, or writing intent but not a full source text

Move directly to Stage 2 only when the user already provides a stable outline or asks specifically for structure refinement.

Move directly to Stage 3 only when:

- the user explicitly says the structure is locked, or
- the user supplies a final outline and asks only for drafting

Even when jumping ahead, rebuild a lightweight internal brief first. Do not draft against a structure you do not understand.

## 2. Stage 1 Checklist

### 2.1 Identify the Requirement

Infer the following from the material and the request:

- mode
- target audience
- reading scenario
- writing purpose
- likely stance
- desired depth
- likely tone

State assumptions when inference is necessary.

### 2.2 Build the Brief

Use `assets/brief-template.md`.

Fill at least:

- creator persona or author position
- audience profile
- writing scenario
- content stance
- writing principles
- core themes
- key viewpoints
- structure requirements

Treat the brief as the baseline for every later decision.

### 2.3 Preprocess the Material

Detect and mark:

- spoken fillers such as "um", "ah", "you know", "就是说", "那个"
- repeated openings or self-corrections
- incomplete clauses that still contain meaning
- digressions that may later become examples instead of main arguments

Do not delete before understanding the role of the passage. Repetition can signal importance.

Skip this subsection when the task is purely topic-driven and there is no source material to clean.

### 2.4 Build the Argument Base

For topic-driven mode, build a minimal argument base before outlining:

- central question
- working thesis
- 3-5 core angles
- likely objections or boundary conditions
- known examples, evidence, or missing evidence

For draft-deepening mode, diagnose:

- the current thesis quality
- thin sections
- missing mechanisms
- unsupported jumps
- duplicated claims

### 2.5 Extract Viewpoint or Argument Units

Break the material into viewpoint units or argument units instead of paragraph-shaped chunks.

Useful unit types:

- claim
- reason
- mechanism
- example
- contrast
- implication

This makes clustering more reliable than operating on raw paragraphs or raw notes.

## 3. Clustering Rules

Merge units when they point to the same thesis even if the wording, example, or confidence level differs.

Keep units separate when they differ in:

- core stance
- causal direction
- abstraction level
- time frame
- speaker perspective

Use the cluster output to preserve nuance while removing duplication.

For topic-driven mode, cluster by argumentative role as well as semantic similarity. Two claims may sound related but still belong to different jobs:

- thesis
- mechanism
- evidence
- counterpoint
- implication

Recommended cluster format:

```text
聚类 1：观点标题
- 原观点：...
- 原观点：...
- 原观点：...
- 合并后观点：...
- 保留的关键细节：...
```

Do not reduce three complementary fragments into a flatter sentence. Merge by integration, not deletion.

## 4. Logic Mapping Rules

Map the strongest visible relations in the source material or the strongest argument path for topic-driven writing.

Use these relation types as defaults:

- causal: because, therefore, leads to, results in
- parallel: first, second, in another dimension, likewise
- progressive: not only, further, deeper, beyond
- contrast: however, unlike, yet, on the contrary
- summary: therefore, in short, overall, in conclusion

When the material is weakly structured, force a minimum chain:

1. premise
2. mechanism
3. consequence
4. implication

If the source or topic does not support a clean chain, say so instead of inventing one.

For topic-driven mode, common high-level structures include:

- question -> answer -> mechanism -> implication
- problem -> diagnosis -> principle -> solution -> limit
- misconception -> correction -> evidence -> conclusion
- observation -> interpretation -> consequence -> action

Choose the structure that best serves the reader, not the most decorative one.

## 5. Stage 2 Checklist

Treat Stage 2 as a structure-locking step.

Do all of the following:

- start from the accepted brief and viewpoint clusters
- keep the macro frame when possible in source-driven mode
- keep the strongest existing frame and repair weak branches in draft-deepening mode
- choose the clearest reader path in topic-driven mode
- merge duplicate branches
- reorder only adjacent or clearly misplaced blocks
- assign each section a main claim
- assign each section supporting points or evidence
- write transitions where logic would otherwise jump

Mark changes explicitly:

- `沿用原结构`
- `合并重复观点`
- `顺序微调`
- `新增过渡`
- `按命题重建结构`

Do not generate the full article here. Keep the output structural.

## 6. Stage 3 Checklist

Write against the approved structure only.

Deepen a section with one or more of these moves:

- add background that helps the reader enter the topic
- explain the mechanism behind a claim
- supply a concrete example
- add contrast or boundary conditions
- surface implications for the reader

Avoid generic expansion such as:

- empty motivational lines
- repeated summary sentences
- vague "with the development of society" filler
- unrelated facts added only to increase length

Prefer precise, readable Chinese over literal transcript residue.

For topic-driven mode, make sure each section advances the thesis instead of circling around it in different words.

## 7. One-Shot Mode

If the user explicitly asks for one-pass output:

- run the three stages internally
- present a compressed Stage 1 and Stage 2 summary before or alongside the final draft when useful
- state key assumptions
- keep the final structure consistent with the internally locked outline

Do not skip the reasoning steps just because the visible response is compressed.

## 8. Common Failure Patterns

### 8.1 Over-expansion

Symptom:

- the draft is longer but not deeper

Fix:

- cut any paragraph that adds no mechanism, example, or implication

### 8.2 Source Drift

Symptom:

- the final article sounds polished but no longer reflects the original viewpoint

Fix:

- compare each section against the brief and the merged viewpoints

### 8.3 Thesis Drift

Symptom:

- the article begins on one question and ends on another

Fix:

- compare each section against the brief and the locked Stage 2 section jobs

### 8.4 Duplicate Claims in Different Sections

Symptom:

- the same idea appears as separate headings or paragraphs

Fix:

- re-cluster the sections and assign one home for each major claim

### 8.5 Logic Jumps

Symptom:

- sections feel individually fine but do not connect

Fix:

- add explicit transition sentences or reorder the local block

### 8.6 Transcript Residue

Symptom:

- the draft still contains spoken-language pacing, half-sentences, or filler transitions

Fix:

- rewrite for written rhythm while preserving the meaning