Write the spec before the code, then check the built thing against it — so agents build what was agreed, not what they invented.
The English summary is being prepared. Showing the original description.
See how to installReview the instructions the agent will receive before you install.
OpenSpec ensures AI agents build what was planned, not what they invented.
project/
└── openspec/
├── specs/ # Source of truth (what IS built)
├── changes/ # Proposals & WIP (what SHOULD change)
├── ideas/ # Quick-capture for later
└── project.md # Project conventions
Create a change proposal before implementing anything significant.
openspec/changes/add-feature-x/
├── proposal.md # What and why
├── tasks.md # Numbered implementation steps
├── design.md # Technical approach
└── specs/ # Delta specs (what will change)
Work through tasks.md sequentially. Keep specs in sync with code.
When complete: openspec archive <change-id>
Specs merge to source of truth. Change moves to archive.
# List all changes
openspec list
# Show specific change details
openspec show <change-id>
# Validate a change
openspec validate <change-id> --strict
# Archive completed change
openspec archive <change-id> --yes
# List specs
openspec spec list --long
proposal.md format:
# Feature Name
## Summary
One paragraph: what this change does.
## Motivation
Why we need this. What problem it solves.
## Scope
- What's included
- What's explicitly NOT included
## Dependencies
- Other changes this depends on
- External requirements
tasks.md format:
# Implementation Tasks
## Phase 1: Foundation
- [ ] Task 1.1: Description
- [ ] Task 1.2: Description
## Phase 2: Core Implementation
- [ ] Task 2.1: Description
...
openspec validate --strict must passYES - Use OpenSpec for:
NO - Skip OpenSpec for:
OpenSpec tracks WHAT gets built. megg tracks WHY decisions were made.
Use both:
Before marking a change complete:
openspec validate --strict passesNo reviews yet. Tell others what you thought.
No login needed. Share how it worked for you, ask a question, or note where it did not run.
No comments yet. Be the first to write one.
Branch and commit conventions that keep an agent's work reviewable — one branch per task, main stays clean, nothing force-pushed.
The English summary is being prepared. Showing the original description.
Look up how Claude Code itself works — hooks, skills, subagents, MCP, plugins, settings — from the official docs, so answers come from the current spec instead of memory.
The English summary is being prepared. Showing the original description.
Design better agents, skills and hooks for Claude Code, checking the current spec instead of guessing at fields.
The English summary is being prepared. Showing the original description.
Run a feature end to end — scout the code, plan it, build it, and keep a session file so a closed terminal doesn't lose the thread.
The English summary is being prepared. Showing the original description.
Close out a session — summarise what changed, capture what was learned, and prepare the commit.
The English summary is being prepared. Showing the original description.
Hand over a spec and walk away — plans, builds, tests and audits on its own, stopping only at a real blocker.
The English summary is being prepared. Showing the original description.