gsd-claw - SKILL.md Agent Skill

name: GSD Claw

description: Spec-driven development with built-in verification for substantial projects. Use when user wants to plan a project, scope a feature, build something with structure, or says "GSD mode", "let's plan", "scope out", "spec-driven". Workflow is Discuss → Plan → Execute → Verify. Based on glittercowboy's GSD system (MIT license). NOT for quick questions or simple tasks.

GSD Claw

A spec-driven development system adapted for OpenClaw. Transforms vague ideas into verified, working implementations through structured workflows.

Based on: glittercowboy/get-shit-done

Original Author: Lex Christopherson (@glittercowboy)

License: MIT

When to Use

Trigger this skill when user says:

"Let's plan [project/feature]"
"I want to build [something]"
"Scope out [project name]"
"GSD mode" or "spec-driven"
"Plan and execute [task]"
Any substantial project that needs structure

NOT for: Quick questions, simple tasks, chat conversations.

Core Philosophy

From the original GSD:

"The complexity is in the system, not in your workflow."

Principles:

Plans ARE prompts — Executable instructions, not documents to interpret
Verification built-in — Every task has acceptance criteria
Fresh context per task — Subagents prevent context rot
Solo developer + AI — No enterprise theater

Workflow Overview


┌─────────────────────────────────────────────────────────────┐

│  1. DISCUSS     Capture decisions before planning           │

│  2. PLAN        Research → Create verified specs            │

│  3. EXECUTE     Work through tasks with deviation rules     │

│  4. VERIFY      Confirm deliverables actually work          │

└─────────────────────────────────────────────────────────────┘

Phase 1: DISCUSS

Purpose: Capture user's vision BEFORE any planning.

Process:

Ask about the goal (what are we building?)
Identify gray areas based on domain:
- Visual features → layout, interactions, empty states
- APIs → response format, error handling
- Content → structure, tone, depth
For each area, ask until user is satisfied
Document decisions in {project}/.gsd/CONTEXT.md

Output Structure:


# Context: [Project/Phase Name]



## Decisions (Locked)

- [User decision 1]

- [User decision 2]



## Agent Discretion (Freedom Areas)

- [Area where agent can choose approach]



## Deferred Ideas (Out of Scope)

- [Not doing this now]

Rule: Never assume. Always ask.

Phase 2: PLAN

Purpose: Create executable task specs with built-in verification.

Research (Optional but Recommended)

Spawn sub-agent to investigate:

Technical approach options
Potential pitfalls
Best practices for the domain

Save to: {project}/.gsd/RESEARCH.md

Create Plans

Each plan = 2-3 tasks maximum (context efficiency).

Task XML Structure:


<task type="auto">

  <name>Task N: Action-oriented name</name>

  <files>src/path/file.ts, src/other/file.ts</files>

  <action>

    What to do, what to avoid and WHY.

    Be specific. No guessing.

  </action>

  <verify>Command or check to prove completion</verify>

  <done>Measurable acceptance criteria</done>

</task>

Task Types:

type="auto" — Agent executes autonomously
type="checkpoint:verify" — User must verify
type="checkpoint:decision" — User must choose

Plan Verification

Before execution, verify plans against:

Tasks are specific and actionable
Each task has verify + done criteria
Scope matches CONTEXT.md decisions
No enterprise bloat

Save plans to: {project}/.gsd/plans/{phase}-{N}-PLAN.md

Phase 3: EXECUTE

Purpose: Work through tasks with deviation handling.

Execution Rules

Read plan as literal instructions
Execute each task in order
Verify done criteria before moving on
Apply deviation rules automatically
Document everything

Deviation Rules

Apply automatically — track for summary:

|------|---------|--------|------------|

Rule 4 (Architectural) Format:


⚠️ Architectural Decision Needed



Current task: [task name]

Discovery: [what prompted this]

Proposed change: [modification]

Why needed: [rationale]

Alternatives: [other approaches]



Proceed? (yes / different approach / defer)

Fresh Context Pattern

For multi-task execution, spawn sub-agents:

Each sub-agent gets fresh 200k context
Main context stays lean for user interaction
Prevents quality degradation

Task Completion

After each task:

Verify done criteria met
Commit changes (if applicable)
Record in summary
Move to next task

Phase 4: VERIFY

Purpose: Confirm deliverables actually work.

Automated Verification

Run verify commands from each task
Check files exist
Run tests if applicable

User Acceptance

Walk user through testable deliverables:

Extract what they should be able to do now
Present one at a time
Get yes/no or issue description
If issues found → create fix plans

If all pass: Mark phase complete.

If issues: Don't debug manually — create fix plan and re-execute.

File Structure


{project}/

└── .gsd/

    ├── PROJECT.md           # Vision, always loaded

    ├── STATE.md             # Current position, decisions, blockers

    ├── REQUIREMENTS.md      # Scoped v1/v2 requirements

    ├── ROADMAP.md           # Phases and progress

    ├── CONTEXT.md           # User decisions from discuss phase

    ├── RESEARCH.md          # Domain research (optional)

    └── plans/

        ├── 01-01-PLAN.md    # Phase 1, Plan 1

        ├── 01-01-SUMMARY.md # Execution results

        └── ...

Templates: Copy from assets/ directory when initializing project files.

Quick Mode

For small tasks that don't need full planning:

Trigger: "Quick: [task description]"

Process:

Skip research and discussion
Create single plan with 1-3 tasks
Execute with deviation rules
Verify and commit

When to use:

Bug fixes
Small features
Config changes
One-off tasks

State Management

STATE.md Template


# Project State



## Current Position

Phase: [N]

Plan: [N of M]

Status: [planning | executing | verifying | blocked]



## Decisions Made

- [Decision 1] — [rationale]

- [Decision 2] — [rationale]



## Blockers

- [Blocker if any]



## Deviations Applied

- [Rule N] [Description] — [resolution]

Update STATE.md after each significant action.

Anti-Patterns (Banned)

From original GSD:

❌ Enterprise theater

Story points, sprint ceremonies
Human time estimates (days/weeks)
RACI matrices, stakeholder syncs

❌ Vague tasks


<!-- BAD -->

<task>

  <name>Add authentication</name>

  <action>Implement auth</action>

  <verify>???</verify>

</task>

❌ Filler language

"Let me...", "I'd be happy to..."
"Great question!", "Excellent!"

✅ Direct, imperative, technical

OpenClaw Adaptations

Original GSD → OpenClaw Version

| GSD | OpenClaw |

|-----|----------|

| /gsd:command slash commands | Natural language triggers |

| Claude Code Task() spawning | sessions_spawn sub-agents |

| .claude/ directory | {project}/.gsd/ directory |

| Bash scripts | PowerShell/cross-platform |

| git commit protocol | Optional, user preference |

Preserved Core:

XML task structure with verify/done
Deviation rules
Fresh context via sub-agents
State management
Plan verification loop

Example Usage

User: "Let's plan the ORO website SEO overhaul"

Agent (DISCUSS):


Starting GSD workflow for: ORO Website SEO Overhaul



Before planning, I need to understand your vision.



1. What's the primary goal?

   a) Rank higher for specific keywords

   b) Improve technical SEO scores

   c) Both

   d) Something else



2. Any pages that are highest priority?

3. Budget constraints for tools/services?

4. Timeline expectations?

After discussion → creates CONTEXT.md

Agent (PLAN):


━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

 GSD ► PLANNING: ORO SEO Overhaul

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━



Spawning research sub-agent...

[Research complete]



Creating plans...

- Plan 1: Technical SEO fixes (3 tasks)

- Plan 2: Content optimization (2 tasks)

- Plan 3: Performance improvements (3 tasks)



Verifying plans against requirements...

✓ All checks passed



Ready to execute?

Credits

This skill is adapted from Get Shit Done (GSD) by Lex Christopherson (@glittercowboy).

Original Repository: https://github.com/glittercowboy/get-shit-done
License: MIT
Philosophy: Solo developer + AI, no enterprise bloat

Thank you Lex for creating an incredibly effective system for building things with AI.

Version History

1.0.0 — Initial OpenClaw adaptation