From 5% to 90%: Tmall/Taobao Overseas AI Coding SOP Boosts Adoption with SDD All‑in‑One

Background and Challenges

In 2025 AI coding tools such as Cursor, Copilot and Claude Code surged in popularity, making "AI writes code" a reality. The Tmall/Taobao overseas technical team quickly adopted these tools across several business domains (growth, recommendation, supply, core infrastructure). However, three concrete pain points emerged:

Requirement ambiguity amplification : vague PRD descriptions that developers can interpret become amplified errors when handed to AI, leading to costly rework.

Design‑less low‑quality loops : skipping design causes inconsistent architecture, style drift and a "generate → modify → regenerate" cycle with unstable acceptance rates.

Non‑transferable team experience : different teams used AI Agents, Skills + Command, or OpenSpec frameworks, preventing knowledge sharing and keeping AI coding at the "personal skill" level.

Why an SOP Is Needed

Tools alone are not the bottleneck; a standardized operating procedure (SOP) that spans the entire development lifecycle is required. Similar to ISO standards in manufacturing, an AI‑coding SOP ensures every developer can "accurately use AI to write code". The SOP brings three core values:

Experience reproducibility : unified output formats and process standards enable cross‑team knowledge sharing.

Knowledge sedimentation : each SDD spec is archived and fed back into a knowledge base for future reuse.

Measurable quality : stage‑gate checks turn coding quality from luck‑based to guaranteed.

SDD All‑in‑One Skill

The SDD All‑in‑One Skill is the core tool that implements the SOP. It transforms requirements into structured design documents ( design.md) and interface‑level coding instructions ( tasks.md), eliminating ambiguity before coding, anchoring generated code to existing codebases, and enforcing stage‑gate quality checks.

Evolution of the Solution

V1 – Vibe Coding + Rules (2025.05‑2025.09) : manual/half‑automatic rule generation; context quality directly determined code quality. Core idea: "Context decides everything" and the "raw‑to‑rule" methodology where AI first learns coding rules from the project and then uses them for generation.

V2 – SDD Command version (2025.10‑2026.02) : introduced OpenSpec framework and multiple local commands, standardizing five SOP stages (intake, clarification, proposal, design, tasks, archive). Produced templated artifacts such as intake.md, proposal.md, design.md, tasks.md.

V3 – SDD All‑in‑One (2026.03‑present) : a single Skill encapsulating all stages, driven by a state‑machine, platform‑agnostic, and supporting breakpoint recovery.

Core Design Principles (Baselines)

No Design, No Code : code generation is blocked until design.md is persisted.

Design is Truth : all decisions are written back to design.md; conflicts are resolved by the design document.

No Approval, No Execute : only an explicit Task Approved token unlocks execution; vague confirmations like ok or continue are rejected.

Architecture and State Machine

The engine consists of three main generators:

Code‑scan engine : extracts project structure, layer architecture, coding conventions (8 items), feature completion status, and relevant code locations without any preset assumptions.

Design.md generator : produces a call‑chain description (e.g., "Entry method → Business method → Domain method → Data access method") and enforces anchoring to real classes, PlantUML activity diagrams, and external API verification.

Tasks.md generator : each task maps to a single interface method, contains zero code, and includes parameter validation, pre‑check, state check, CAS, transaction boundary, and exception handling.

After generation, a four‑dimensional cross‑validation checks intake↔design coverage, design↔tasks consistency, artifact↔code reality, and format compliance. Validation failures are listed for manual correction.

Workflow Modes

Quick Mode (single‑sentence, clear scope) is triggered by quick_design or >>> and follows a linear path: Init → Requirement → Code Scan → design.md → tasks.md → Task Approved → execution.

Full Mode (complex, full PRD) is triggered by full_design or FULL. It adds an explicit clarification step (Q&A) before intake creation, allowing richer requirement capture and higher accuracy.

Breakpoint Recovery

The state machine records each artifact on disk. When a user sends continue, the Skill reads the latest state and resumes without restarting or skipping any gate. Recovery scenarios range from no artifact (prompt for intent) to partially completed intake.md, design.md, or tasks.md.

Case Studies

Case 1 – High‑Commission Settlement Migration (Full Mode) : The Skill generated an intake.md with 8 "to‑do" items, 3 "not‑do" items, and detailed transaction and gray‑release plans. The clarified design prevented direction‑wrong rework and achieved >90% acceptance.

Case 2 – Distribution Purchase Upgrade (Full Mode) : Compared with a manual solution, the Skill delivered the same functionality in one day instead of 2‑3 days, with a design‑backed guarantee and no need for continuous Agent interaction.

Case 3 – Breakpoint Recovery & Issue Reporting : After an interruption during design generation, sending continue restored the workflow. An issue report via sb-sdd: … automatically created a sub‑task in Aone without affecting the current design.

Adoption Metrics

From May 2025 to the present, AI coding usage grew in a stair‑step fashion:

V1 (2025.07‑2025.10): 5% → 30% (6× increase in 4 months).

V2 (2025.10‑2026.02): 30% → 65% (doubling in 3 months).

V3 (2026.02‑now): 65% → 90% (continuous climb).

These numbers are visualised in the internal chart (image omitted for brevity).

Cultural Practices Supporting AI Coding

Cross‑team sharing : regular demos showing "how to finish a real requirement with AI".

Data‑driven coaching : analysis of individual AI‑coding ratios to provide targeted assistance.

Pair‑with‑AI : senior engineers mentor newcomers in providing effective prompts, reviewing AI output, and correcting mistakes.

Bi‑weekly retrospectives : review successful and failed cases to iteratively improve the SOP.

Installation Quick‑Start

One‑click installation via the Aone Copilot skill store:

# Aone Copilot: search "sdd-all-in-one"
# Qoder: a1 skill install sdd-all-in-one --agent qoder
# Claude Code: a1 skill install sdd-all-in-one --agent claude

First run automatically installs required MCPs (code, Yuque, coop) and requires an IDE restart.

The SDD trace tool must also be installed:

npx -y @ali/ai-coding-trace-lite@latest setup

Future Roadmap

Short‑term plans include a full‑stack AI skill that auto‑detects front‑end, back‑end or mobile contexts and routes to the appropriate MCP, achieving true AI full‑stack development. Additionally, three test‑node slots (A, B, C) will embed test‑point generation, cross‑validation, and automated test‑case creation, making "file is contract" the core guarantee.

Long‑term vision expands the current five‑stage SOP (requirement → design → coding → review → test) to a six‑stage lifecycle covering release and operations, under a three‑layer Spec / Harness / Platform architecture that continuously learns from AI mistakes and solidifies rules, skills, and platform support.