# SpekLess: A Lightweight Spec-First Development Workflow for AI Coding Agents

> SpekLess is a lightweight spec-first development workflow designed for AI coding agents (e.g., Claude Code). It uses a single design document (spec.md) and an append-only work log (execution.md), along with ten slash command skills, to implement the full lifecycle of clarification → planning → implementation → verification—without the need for state machines, lock files, or checkpoint files.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-04-09T20:10:26.000Z
- 最近活动: 2026-04-09T20:26:51.866Z
- 热度: 157.7
- 关键词: AI 编码, Claude Code, 规格驱动, 开发工作流, 斜杠命令, 文档即状态, Agent 工作流
- 页面链接: https://www.zingnex.cn/en/forum/thread/spekless-ai-agent
- Canonical: https://www.zingnex.cn/forum/thread/spekless-ai-agent
- Markdown 来源: floors_fallback

---

## Introduction / Main Post: SpekLess: A Lightweight Spec-First Development Workflow for AI Coding Agents

SpekLess is a lightweight spec-first development workflow designed for AI coding agents (e.g., Claude Code). It uses a single design document (spec.md) and an append-only work log (execution.md), along with ten slash command skills, to implement the full lifecycle of clarification → planning → implementation → verification—without the need for state machines, lock files, or checkpoint files.

## Why Do We Need SpekLess?

Spec-driven development in AI coding tools like Claude Code is valuable, but existing tools have friction:

## Problems with GSD (GitHub Spec-Driven)

- **High token consumption**: Each workflow step (researcher, planner, plan-checker, executor, verifier, integration-checker, nyquist-auditor...) starts a cold-booted sub-agent that re-reads files and writes its own artifacts
- **Forced atomic commits**: Pollutes git history and locks you into a rigid state machine

## Problems with GitHub SpecKit

- Scatters a single feature across three files (`spec.md`, `plan.md`, `tasks.md`)
- Uses a template-driven process that doesn't integrate with Claude Code skills

## Problems with Pure ADR/RFC Documents

- Human-readable, but agents can't get help—each discussion, planning, and verification is ad-hoc

## SpekLess's Solution

SpekLess is a minimal system that keeps the good parts and removes the rest:

## Core Design Principles

1. **One living `spec.md` per feature**
   - Reads like an RFC
   - Includes Context, Discussion, Plan, and Verification sections
   - All content for a feature is in one place

2. **Append-only `execution.md` work log**
   - Human-readable narrative of actual work
   - Coexists with the spec
   - Replaces atomic commit discipline and doesn't force commits at others' paces

3. **Flat feature list**
   - No nested hierarchy of milestones → phases → tasks
   - Large work is broken into sibling documents via the `part_of:` frontmatter field

4. **Ten skills instead of thirty sub-agents**
   - A single main agent drives everything via slash commands
   - Sub-agents are only used as context firewalls for broad codebase reads, never as pipeline steps

5. **Intervene at any time**
   - Documents are the state
   - No STATE.md, no checkpoint files, no locked step order
   - Edit any section, re-run any skill, at any time

## Entry Points (Choose One at Startup)

| Skill | Use Case |
|-------|----------|
| `/spek:kickoff` | Launch a greenfield project. Runs an extended PRD-style discussion, writes to `.specs/project.md`, and proposes setting up initial feature folders |
| `/spek:new` | Add a new feature to any project. Creates a skeleton feature folder—nothing more |
| `/spek:adopt` | Retroactively document existing code. Reads actual files (via the Explore sub-agent), reverse-engineers Context and Plan sections, and marks all tasks as completed. Exclusive to SpekLess |