Vibe Flow is a hackable set of prompts that enables Spec-Driven Development for professional developers.
It works well with any agent powered by a coding model such as:
- Claude Opus 4+ or Claude Sonnet 4+ (Anthropic)
- GPT 5+ (OpenAI)
- Composer 1 (Cursor)
- Ensure your agent uses a capable coding model.
- Give it this installation prompt.
It will install the following structure and plug your AI instructions file into it:
_docs/
├── vibe-flow/
│ ├── Discuss-Then-Do Protocol.md
│ ├── How to Write a Technical Specification.md
│ ├── How to Write Implementation Plans.md
│ └── Vibe Flow Guide.md
├── Code Style Guidelines.md
├── How to Write Unit Tests.md
└── Writing Documentation.md
_plans/
AGENTS.md
Then, start using the workflow.
A specification can be written long before the implementation. The agent helps you write it by investigating and initiating a discussion:
Read your documentation, then help me write a new spec.
[something to do]Or, if you use Claude Code or Cursor, you can use the command: /spec [something to do].
The agent will discuss with you, then write a _plans/123/A1-spec.md file.
Note: 123 is the ticket ID. If it can be deduced from the branch name, it will. Otherwise the agent will ask you. A1 means it's the first file of cycle A (files are organized by cycles).
Plans orchestrate what agents or subagents will do:
Read your documentation, then write plans.Or, if you use Claude Code or Cursor, you can use the command: /plan.
The agent reads the spec and writes plan(s) in _plans/123/A2-plan*.md.
Clear the context, then execute the plan(s):
Execute the plan `_plans/123/A2-plan-orchestrator.md`The agent executes and writes handover document(s) (.summary.md files).
There is also a lighter prompt for small tasks without spec/plans. Here's how to trigger it:
Read your documentation first. I need a DTDP.
[something to do]Or, if you use Claude Code or Cursor, you can use the command: /dtdp [something to do].
The agent will discuss first, then it will directly work on the codebase. At the end a _plans/123/A1-summary.md file will be written.
The documentation for AI agents must be reorganized into multiple files, because:
- We want our technical documentation to be sustainable and not specific to how agents work today.
- We want the same documentation to be shared among several agents and humans.
- We don't want to fill the context window with unnecessary instructions.
Everything must be written in (git-ignored) local files, because:
- The context window is limited, the compression mechanism is opaque, and we want to be able to continue an unfinished task in a fresh session.
- It's a way to keep track of what we agreed with the agent and what has been done.
There are no guidelines. This is just a starting point for your own workflow. Feel free to adapt it to your needs.
See the ParoiCMS repository for a real-world example with its _docs/ directory.
Two migration prompts are available:
- Upgrade from AI Workflow: replace your
_docs/ai-workflow/directory and the instruction file with the Vibe Flow directory andAGENTS.md. - Update Vibe Flow: overwrite your
_docs/vibe-flow/directory with the latest versions of the Vibe Flow prompts, and reference them in yourAGENTS.md.