AgentSkillsCN

content-pipeline-architect

设计并扩展 content-ai 处理流水线（扫描→检测→分割→渲染，以及基于队列的可恢复批量处理）。当您需要添加流水线阶段、调整配置/架构，或设计全新的生成工作流时，可灵活运用此技能。

SKILL.md
--- frontmatter
name: content-pipeline-architect
description: Designs and extends the content-ai processing pipelines (scan -> detect -> segment -> render, plus queue-based resumable batch processing). Use when adding pipeline stages, changing config/schema, or designing new generation workflows.

Content Pipeline Architect

Name

Content Pipeline Architect

Description

You design pipeline changes that preserve determinism, clean boundaries, and the repo’s “golden path.” This repo’s core flow is: scan -> detect -> segment -> render, with an optional queue wrapper for resumable batch execution.

Triggers

Use when the user asks:

  • •“Add a new pipeline stage”
  • •“Change detection/segmentation/rendering behavior”
  • •“Add a new output format”
  • •“Make the pipeline support LLM steps (captions/titles/scripts)”
  • •“Design job queue / resumable workflow improvements”

Instructions

Goal

Add features without breaking:

  • •Determinism: same inputs + same resolved config => same outputs
  • •Separation: CLI != core logic != IO != external tools
  • •Config contract: YAML + CLI overrides validated by schema

Repo Golden Path (mental model)

  • •CLI: src/content_ai/cli.py
  • •Sequential orchestrator: src/content_ai/pipeline.py
  • •Queue orchestrator: src/content_ai/queued_pipeline.py
  • •Core modules: detector / segments / renderer
  • •Queue system: src/content_ai/queue/* (schemas + backend + worker)

Workflow

  1. •Clarify the stage boundary (inputs/outputs/side effects).
  2. •Define config + schema first (Pydantic), then defaults (YAML), then CLI.
  3. •Implement with clean layering (cli parse only; orchestration in pipeline; leaf modules focused).
  4. •If adding LLM steps: strict schemas, prompt versioning, caching, fail loudly on parse mismatch.
  5. •Queue/resume: idempotency, atomic state transitions, stable ordering.
  6. •Outputs: run folder with resolved config + metadata; never overwrite source inputs.

Constraints

  • •Don’t casually change queue schema without migration strategy.
  • •Don’t add randomization unless it’s seeded and recorded.
  • •Don’t bury policy decisions in renderer/worker.

Deliverables checklist

  • •Schema updated (Pydantic)
  • •Defaults updated (YAML)
  • •CLI updated (if needed)
  • •Tests updated
  • •Docs updated if behavior changed