Flow

Lightweight skill for AI automation flows.

Set of commands to structure and drive the cycle of analysis, review, implementation and validation of a request or task.

Installation

Bash

curl -fsSL https://github.com/RivoLink/flow/raw/main/scripts/install.sh | bash

PowerShell

irm https://github.com/RivoLink/flow/raw/main/scripts/install.ps1 | iex

The script will prompt interactively for the harness (Claude, Codex, or Gemini).

Advanced usage

To pass options, download the script first then run it with arguments.

Bash

curl -fsSL https://github.com/RivoLink/flow/raw/main/scripts/install.sh -o flow.sh
bash flow.sh --claude --user --as-command --fr

PowerShell

irm https://github.com/RivoLink/flow/raw/main/scripts/install.ps1 -OutFile flow.ps1
./flow.ps1 -Claude -User -AsCommand -Fr

Options

Flag	Description	Default
claude | codex | gemini	Target harness	Interactive prompt
user | project	Installation scope	project
as-skill | as-command	Installation mode	as-skill
en | fr	Language for .md files	en

Features

Command	Description
flow help	Display available commands and workflow
flow init	Initialize the project working environment
flow best-practices	Add best practices to the project reference
flow asking	Create a request file from a prompt
flow analysis	Analyze a request without modifying code
flow review	Critical review of the analysis
flow process	Apply code modifications
flow validate	Critical review of code modifications
flow re-process	Fix code according to the validation report

Workflow

The cycle starts with asking and ends when validate returns Valid.

asking ──> analysis <──> review ──> process ──> validate <──> re-process

• Loop 1: analysis and review until the analysis is consistent.
• Loop 2: validate and re-process until the status is "Valid".

How it works

Flow drives a task through a fixed cycle. Each step reads and writes files under .flow/, so progress is always inspectable on disk.

Artifacts

• .flow/asks/<slug>.txt : the reformulated request.
• .flow/outs/<slug>-analysis.md : the analysis. A Review Report section and then a Validation Report section are appended in place over the cycle.
• .flow/docs/best-practices.md : project conventions read by process and re-process.

Two feedback loops

• Loop 1 : analysis ↔ review. Iterate until the analysis is coherent and every review question is answered.
• Loop 2 : validate ↔ re-process. Iterate until the validation status is Valid.

Exit

The cycle ends when validate returns Valid.

Commands

Workflow commands first, then setup and meta.

flow asking <prompt>

• Purpose : Turn a free-form prompt into a clean request file.
• Input : Prompt text.
• Output : .flow/asks/<slug>.txt.

flow analysis <path>

• Purpose : Produce a detailed analysis of the request. No code change.
• Input : .flow/asks/<slug>.txt.
• Output : .flow/outs/<slug>-analysis.md.

flow review <path>

• Purpose : Critique the analysis for inconsistencies, questions and risks.
• Input : Analysis path.
• Output : Review Report section appended to the analysis.

flow process <path>

• Purpose : Apply code modifications by following the analysis.
• Input : Analysis path.
• Output : Source code modifications.

flow validate <path>

• Purpose : Critique the code modifications against the analysis.
• Input : Analysis path.
• Output : Validation Report section appended to the analysis, with status Valid or Not valid.

flow re-process <path>

• Purpose : Fix the points raised by the last validation report.
• Input : Analysis path with a Validation Report.
• Output : Source code fixes.

flow init

• Purpose : Initialize the project working environment.
• Input : None.
• Output : .gitignore entry and the .flow/ directory tree.

flow best-practices <comment>

• Purpose : Record project conventions used by process and re-process.
• Input : Comment text.
• Output : Appended to .flow/docs/best-practices.md.

flow help

• Purpose : Print the command list and the workflow diagram.
• Input : None.
• Output : Console message.

Languages

Flow is available in English (en, default) and French (fr).

License

MIT