How Specflow Turns AI Coding into a Structured Development Pipeline

Background

The Tianji frontend team encountered inefficiencies when using AI‑assisted coding for complex backend forms and data transformations. The root causes were context gaps and lack of shared requirement understanding . To address this, the team adopted a Spec‑Driven Development (SDD) approach that requires precise, machine‑readable specifications before code generation.

Solution Research

OpenSpec – Lightweight, Change‑Oriented Agile Tool

Core point: Promotes atomic changes through a Proposal‑Apply loop, breaking large features into small change packages.

GitHub Spec Kit – Industrial‑Grade Collaboration Protocol

Core point: Enforces "write specs before tasks" via a Constitution that defines technical baselines, ensuring AI‑generated code follows project standards.

Insight: Provides a gate‑keeping mechanism; if the specification stage is misaligned, subsequent implementation diverges.

BMAD‑METHOD – Multi‑Agent Collaborative Engineering Framework

Core point: Treats AI as a full expert team (PM, architect, QA) rather than a simple assistant.

Insight: Introduces "role‑thinking isolation" so the Plan stage behaves like an architect instead of a developer writing code on the fly.

Why Combine the Three?

Existing IDE extensions (e.g., Cursor) still required frequent file switching and suffered from loss of conversational state. To reduce mental load and preserve context, the team built Specflow , a custom AI development flow that merges OpenSpec’s lightweight nature, Spec Kit’s rigor, and BMAD’s multi‑role thinking.

Solution Design – Reshaping the AI Development Pipeline

Specflow implements a Spec‑Driven Development workflow consisting of four core phases, each triggered by a single command ( /specflow).

Overall Architecture

The design is distilled into four dimensions:

Full‑link closed loop: From vague requirements to deliverable code.

Strict physical gatekeeping: Prevents progression with incomplete specifications.

Single‑command state machine: Zero‑cognitive‑load automation.

SSOT & role isolation: All artifacts stored in a single source of truth ( plan.md) and AI agents operate with isolated prompts.

Full‑Link Closed Loop

Specify (Problem Clarification): After product documentation is confirmed, AI scans the codebase, identifies logical gaps, and produces a complete business spec.

Plan (Technical Modeling): Maps the business spec to technical paths, defines interfaces and task breakdowns, ensuring every line of code is traceable.

Implement (Atomic Implementation): Executes tasks in groups with a breakpoint Review mechanism, allowing human‑in‑the‑loop verification.

Strict Physical Gatekeeping

Logical Blocker: If the Specify stage lacks detail or the Plan document contains unresolved Block items, the workflow halts.

Deterministic Assurance: Forces developers to think and write clearly before coding, eliminating rework caused by misunderstandings.

Single‑Command State Machine

Physical tracking: Users run /specflow and the system automatically detects the current mode (PM, architect, developer) by inspecting local files such as plan.md.

Automatic stage detection: No manual tool switching; the AI adapts its role based on file status.

SSOT and Role‑Thinking Isolation

All artifacts are stored in a Single Source of Truth ( plan.md), allowing AI to reference a single document for context, reducing token consumption and preserving consistency. Multiple agents are assigned specific prompts for each stage (PM, TL, Dev, Admin), providing internal audit and calibration to mitigate individual bias.

Implementation Details

Standard Workflow Commands

Specify: /specflow-specify – Senior PM defines "what" and "why"; output to ai-docs/{ID}/specify.md.

Plan: /specflow-plan – Architect creates functional contracts and Phase‑3 execution paths; output to ai-docs/{ID}/plan.md.

Implement: /specflow-implement – Software engineer performs atomic coding with real‑time sync and logs.

Archive: /specflow-archive – Knowledge manager generates a summary, moves docs to history/{year}/{quarter}/, and updates ARCHIVE_SUMMARY.md.

Installation & Configuration

IDE: Only supports Cursor.

Node.js: >=18.0.0.

Install the CLI from the private iQIYI mirror and run it globally.

Initialize the project root to inject commands and templates into the .cursor folder.

Core Features Overview

Commands: Predefined AI command set to improve code generation quality.

Templates: Standardized code templates ensuring consistent team style.

Future Roadmap

Specflow 2.0 will introduce Subagents for deeper role isolation and context trimming, as well as Agent Skills that trigger automatically based on file changes or intent, moving from manual /specflow invocations to autonomous workflow execution.

Conclusion

AI‑assisted coding should prioritize precision over speed. Specflow shifts the focus from the "coding‑debug" phase to the "requirement‑design" phase, making code generation cheap and reliable while the real competitive edge lies in defining clear requirements, decomposing tasks, and enforcing standards.