sdd

SkillDocs & knowledge

Sdd gives your AI a set of Specification Driven Development workflow commands and agents, based on Github Spec Kit and OpenSpec. Once added, your AI can run development work in a spec-driven way, with specialized agents keeping context organized and checking the quality of the results.

Then ask your AI: use the sdd skill

Details

Instructions available. Your AI can read the instructions. Execution depends on the setup they require.

After adding the pack, ask your AI to run its Specification Driven Development commands to start a task in a spec-driven way.

What your AI can do with it

  • Run Specification Driven Development workflows
  • Follow the Github Spec Kit and OpenSpec approach
  • Use ready-made workflow commands to structure tasks
  • Manage context with specialized agents
  • Review work quality through dedicated agents

What this skill tells your AI

The instructions your AI receives, as published by asiaostrich/universal-dev-standards in skills/spec-driven-dev/SKILL.md and read by Ahel’s review.

Spec-Driven Development Assistant | 規格驅動開發助手

Create, review, and manage specification documents before writing code.

在撰寫程式碼前,建立、審查和管理規格文件。

When to Use /sdd vs uds spec | 何時使用

Scenario/sdduds spec
Formal feature development with review cycle✅❌
Full spec lifecycle (Draft → Archived)✅❌
Quick prototyping / Vibe coding❌✅
Small incremental changes❌✅
Stakeholder sign-off required✅❌
Micro-spec from natural language intent❌✅

/sdd = Full specification lifecycle for formal development uds spec = Lightweight micro-specs for rapid iteration

/sdd = 正式開發的完整規格生命週期 uds spec = 快速迭代的輕量微規格

TL;DR Quick Checklist | 快速檢查清單

  • Search existing specs: look in specs/, docs/specs/, or project spec directory
  • Decide scope: new feature vs modify existing capability
  • Pick a unique spec ID: SPEC-NNN or kebab-case change ID
  • Write proposal with clear AC (Given/When/Then format)
  • Get approval before implementation begins
  • Implement tasks sequentially, verify against spec
  • Archive spec after completion

Decision Tree | 決策樹

New request? | 新需求?
├─ Bug fix restoring spec behavior? → Fix directly | 直接修復
├─ Typo/format/comment? → Fix directly | 直接修復
├─ Dependency update (non-breaking)? → Fix directly | 直接修復
├─ New feature/capability? → Create proposal | 建立提案
├─ Breaking change? → Create proposal | 建立提案
├─ Architecture change? → Create proposal | 建立提案
├─ Agent/role definition (spans multiple features)? → Use spec-type: agent | 使用 Agent SPEC template
└─ Unclear? → Create proposal (safer) | 建立提案(較安全)

Workflow | 工作流程

DISCUSS ──► CREATE ──► REVIEW ──► APPROVE ──► IMPLEMENT ──► VERIFY ──► ARCHIVE

0. Discuss - Clarify Scope | 釐清範圍

Capture gray areas, establish governing principles, resolve ambiguities before writing spec.

1. Create - Write Spec | 撰寫規格

Define requirements, technical design, acceptance criteria, and test plan.

2. Review - Validate | 審查驗證

Check for completeness, consistency, and feasibility with stakeholders.

3. Approve - Sign Off | 核准

Get stakeholder sign-off before implementation begins.

4. Implement - Code | 實作

Develop following the approved spec, referencing requirements and AC.

5. Verify - Confirm | 驗證

Ensure implementation matches spec, all tests pass, AC satisfied.

6. Archive - Close | 歸檔

Archive completed spec with links to commits/PRs.

Spec States | 規格狀態

StateDescription說明
DraftWork in progress草稿中
ReviewUnder review審查中
ApprovedReady for implementation已核准
ImplementedCode complete已實作
ArchivedCompleted or deprecated已歸檔

Spec Structure | 規格結構

# [SPEC-ID] Feature: [Name]

## Overview
Brief description of the proposed change.

## Motivation
Why is this change needed? What problem does it solve?

## Requirements
### Requirement: [Name]
The system SHALL [behavior description].

#### Scenario: [Success case]
- **GIVEN** [initial context]
- **WHEN** [action performed]
- **THEN** [expected result]

## Acceptance Criteria
- AC-1: Given [context], when [action], then [result]

## Technical Design
[Architecture, API changes, database changes]

## Test Plan
- [ ] Unit tests for [component]
- [ ] Integration tests for [flow]

Agent SPEC Structure | Agent 規格結構(spec-type: agent)

# [SPEC-ID] Agent: [Role Name]
<!-- spec-type: agent -->
<!-- agent-id auto-referenced by feature SPECs -->

## Role Definition
- **Role**: [Agent Name]
- **Responsibility**: [One sentence]
- **Autonomy Level**: L[1-5] (per DEC-065)

## Capability Scope
**Owns:**
- [Capability 1]
- [Capability 2]

**Does NOT own:**
- [Explicit exclusion]

## Interface Contract
### Input
| Message Type | Required Fields | Optional Fields |
|---|---|---|
| [Type] | [fields] | [fields] |

### Output
| Artifact Type | Success Condition | Failure Condition |
|---|---|---|
| [Type] | [condition] | [condition] |

## Agent Interactions
- **Upstream**: [Who calls this agent]
- **Downstream**: [Who this agent calls]
- **Parallel**: [Agents working alongside]

## Related Feature SPECs
- [SPEC-NNN] — [This agent's role in that spec]

Scenario Formatting Rules | 場景格式規則

  • Use #### Scenario: (h4 header) for each scenario
  • Every requirement MUST have at least one scenario
  • Use GIVEN/WHEN/THEN format for structured behavior
  • Use SHALL/MUST for normative requirements, SHOULD for recommendations

Delta Operations | 變更操作

When modifying existing specs, use delta sections:

OperationDescription說明
## ADDED RequirementsNew capabilities新增功能
## MODIFIED RequirementsChanged behavior修改行為
## REMOVED RequirementsDeprecated features移除功能
## RENAMED RequirementsName changes重新命名

Cross-Artifact Analysis (/sdd analyze) | 跨 artifact 一致性檢查

The executable face of the acceptance-criteria-traceability standard + forward-derivation single-spine principle (XSPEC-262). Validates that every test is a faithful projection of the AC spine across specs. 本命令是 acceptance-criteria-traceability + forward-derivation single-spine 的可執行面,驗證每個測試是否忠實投影 AC 主幹。

SignalMeaningGate
orphan test@SPEC-NNN @AC-N references an AC no spec defines / 引用不存在的 AC🔴 BLOCKING
uncoveredan AC has no @SPEC-NNN @AC-N reference / AC 無測試引用report only / 僅報告
not_implementedAC marked so in its .ac.yaml / .ac.yaml 標記🔴 BLOCKING before UAT
cross-spec conflictsame AC id defined in >1 spec / 同 AC id 跨多 spec🔴 BLOCKING
orphan .featureGherkin @AC-N tag referencing a non-existent AC / @AC-N 引用不存在🔴 BLOCKING
AC w/o scenarioAC has no .feature scenario (when BDD in use) / AC 無 BDD scenarioreport only / 僅報告
user-guide driftuser-guide T-N with no matching journey/E2E test id / 手冊 T-N 無對應測試🔴 BLOCKING

Coverage % uses the acceptance-criteria-traceability formula (not_implemented excluded). --json for CI.

npm run sdd:analyze -- --specs specs --tests tests [--userguide docs] [--json]

--userguide <dir> enables user-guide↔E2E drift detection (T-NNN, XSPEC-260/257). / 啟用手冊↔E2E drift 偵測。

vs /ac-coverage: ac-coverage = per-spec detailed AC↔test matrix;/sdd analyze = cross-spec/batch consistency + orphan detection(互補、不取代)。

Spec-vs-Code Convergence Check | 規格與實作漂移偵測(2026-07-10,spec-kit /converge 借鑑)

vs /sdd analyze:/sdd analyze 問「測試有沒有正確引用 AC」(引用完整性,deterministic script);convergence check 問「程式碼的實際行為是否真的符合 spec 描述」(語意漂移,需要讀碼判斷,走 sdd.flow.yaml verify 階段的 spec-match-check ai-check step)——兩者互補、檢查的是不同層次的問題,不重疊。

雖然定義在 7-phase 狀態機的 verify 階段(一次性 gate),這個檢查可以在任何時間點對已歸檔的 spec 獨立重跑——用於偵測 spec 歸檔後、code 持續演進累積出的漂移(例如後續 commit 改了行為但沒回頭更新 spec)。重跑時:

  • 落差分類:missing(spec 要求但沒做)/partial(做一半)/contradicts(行為與描述不同)/unrequested(做了沒被要求的事)
  • 嚴重度:CRITICAL(違反 spec 明確 MUST)/HIGH/MEDIUM/LOW
  • 唯讀比對——只列出發現,不自動修改 spec 或程式碼
  • 適合定期(如季度)對重要 spec 抽查,或懷疑特定 spec 已過時時針對性執行

Usage | 使用方式

/sdd                     - Interactive spec creation wizard | 互動式規格建立精靈
/sdd auth-flow           - Create spec for specific feature | 為特定功能建立規格
/sdd review              - Review existing specs | 審查現有規格
/sdd analyze             - Cross-artifact consistency check | 跨 artifact 一致性檢查
/sdd --sync-check        - Check sync status | 檢查同步狀態

Next Steps Guidance | 下一步引導

After /sdd completes, the AI assistant should suggest:

規格文件已建立。建議下一步 / Specification document created. Suggested next steps:

  • 執行 /derive 從規格推導測試工件 ⭐ Recommended / 推薦 — Derive test artifacts from spec
  • 執行 /derive bdd 僅推導 BDD 場景 — Derive BDD scenarios only
  • 執行 /derive tdd 僅推導 TDD 骨架 — Derive TDD skeletons only
  • 審查 AC 完整性,確保所有驗收條件可測試 — Review AC completeness
  • 檢查 UDS 規範覆蓋率 → 執行 /audit --patterns — Check UDS standard coverage → Run /audit --patterns

Reference | 參考

AI Agent Behavior | AI 代理行為

完整的 AI 行為定義請參閱對應的命令文件:/sdd

For complete AI agent behavior definition, see the corresponding command file: /sdd

Signals

GitHub stars
76
Forks
13
Last commit
Oct 2026
Advanced
Item type
skill
Key
sdd
Source
github.com/asiaostrich/universal-dev-standards