# How User Confirmation Is Handled Between Pipeline Stages in Cangjie-Skill

> Discover how Cangjie-Skill ensures user confirmation between pipeline stages using automated self-checks and algorithmic cross-validation for reproducible workflows.

- Repository: [kangarooking/cangjie-skill](https://github.com/kangarooking/cangjie-skill)
- Tags: how-to-guide
- Published: 2026-07-22

---

**User confirmation between pipeline stages in Cangjie-Skill is handled through automated self-checks and algorithmic cross-validation rather than interactive prompts, ensuring a fully reproducible workflow without manual intervention.**

The Cangjie-Skill project implements a sophisticated document processing pipeline that eliminates traditional manual approval gates. According to the `kangarooking/cangjie-skill` repository, user confirmation between pipeline stages is replaced by systematic automated verification embedded directly into the workflow logic. This design ensures that extracted knowledge units flow seamlessly from extraction to delivery without requiring human intervention at each transition point.

## Stage 1: Automated Self-Checks Replace Manual Approval

The pipeline initiates with five parallel sub-agents extracting candidates simultaneously. Rather than pausing for human review, each agent performs an **automated self-check** (输出前的自检) before committing any output. This verification happens entirely within the extraction scripts through simple conditional logic.

### The Three Pre-Output Verification Questions

Each extractor asks itself three internal questions implemented as `if` checks in the extraction logic:

1. Does this unit have a clear source in the book?
2. Is it within the responsibility scope of this extractor?
3. Has it already been extracted by another extractor?

Only when all three answers are affirmative does the candidate get written to `candidates/<type>.md`. This automated gatekeeping ensures data quality without introducing human bottlenecks.

```yaml

# Example output structure after passing the three-question verification

id: f01
title: 逆向思维
type: framework
source_chapter: 第三讲
source_quote: |
  "反过来想,总是反过来想..."
summary: |
  …
tags: [decision, mental-model]

```

## Stage 1.5: Triple-Verify Cross-Validation

Upon completion of Stage 1, results flow directly into **Stage 1.5** (the "triple-verify" stage) without human intervention. As documented in [`methodology/03-stage1.5-triple-verify.md`](https://github.com/kangarooking/cangjie-skill/blob/main/methodology/03-stage1.5-triple-verify.md), this stage performs a **three-way cross-validation** across the different extractor outputs.

### Algorithmic Conflict Resolution

This validation layer merges overlapping candidates and resolves conflicts through deterministic algorithms. The process is entirely programmatic—no human must press "Continue" or "Approve" to proceed. The system evaluates consistency across the parallel extraction outputs and consolidates them into a unified candidate set based on predefined merge rules.

## Subsequent Stages: Deterministic Assertions

Later stages—including the "pressure-test" and "deliver" phases—continue the same automated pattern. These stages run deterministic scripts that consume the markdown files produced by previous steps. The pipeline only halts execution if a **programmatic assertion fails**, such as a missing required field or a schema validation error. According to the source code in [`methodology/02-stage1-parallel-extract.md`](https://github.com/kangarooking/cangjie-skill/blob/main/methodology/02-stage1-parallel-extract.md), this architecture keeps the workflow fully reproducible and eliminates idle time waiting for manual approvals.

## Summary

- **No interactive pauses**: The Cangjie-Skill pipeline never waits for human confirmation between stages.
- **Automated verification**: Each stage embeds self-checks (three questions for extractors, three-way validation for mergers) that gate progress algorithmically.
- **Deterministic flow**: Stage transitions depend solely on programmatic assertions and file-based handoffs, not manual approval clicks.
- **Reproducible design**: By removing human decision points from the inter-stage transitions, the workflow ensures identical outputs across runs.

## Frequently Asked Questions

### Does Cangjie-Skill require manual approval between pipeline stages?

No. The system replaces manual approval with automated verification steps. Each stage performs self-checks—such as the three-question validation in Stage 1 and the three-way cross-validation in Stage 1.5—before automatically passing results to the next stage.

### What triggers a pipeline stage to proceed to the next?

Stage transition is triggered by successful completion of programmatic assertions. For extraction agents, this means answering "yes" to all three internal questions (source clarity, scope alignment, and uniqueness). For the triple-verify stage, this means successfully merging overlapping candidates without conflicts.

### How does the system prevent duplicate extractions without human review?

Each extractor checks against existing candidates by asking "Has it already been extracted by another extractor?" before writing to `candidates/<type>.md`. Additionally, Stage 1.5 performs algorithmic deduplication through three-way cross-validation across all parallel extractor outputs.

### What happens if an automated check fails during stage transition?

If any verification check fails—such as a candidate failing the three-question test or a validation error in the triple-verify stage—the pipeline halts at that specific assertion point. This failure is reported programmatically rather than being queued for human confirmation, allowing for immediate debugging or automated retry logic.