LAISI -- Let AI Supervise Itself

A workflow harness for AI-generated content. Define steps in YAML, validate every output against XSD schemas, and let the CLI handle retries. The LLM produces content; LAISI makes sure it's correct.

Installation

git clone https://github.com/twissmueller/laisi.git
cd laisi
npm install
npm run build
npm link

This makes the laisi command available globally.

Quickstart

mkdir my-blog && cd my-blog
laisi init --workflow blog-post
laisi run --all
cat .laisi/review.xml

This runs the built-in blog-post workflow: three steps (outline, draft, review), each producing validated XML. The whole pipeline takes about a minute.

How It Works

• Workflows are YAML. Each workflow is a directory with a workflow.yml, XSD schemas, and prompt templates. No code to write.
• Steps chain together. A step declares a predecessor -- its output XML becomes input context for the next step.
• Each step: optional pre_script -> LLM call (prompt + predecessor XML) -> validate output against XSD -> optional post_script.
• One step per run. laisi run executes one step and exits. laisi run --all runs all remaining steps, stopping on failure.
• Files are state. Completed steps produce <step-id>.xml in .laisi/. The CLI checks what exists and picks up where it left off.
• Failed steps get a .failed marker. Delete it to retry.

Creating Your Own Workflow

A workflow is a directory with a naming convention:

workflows/my-workflow/
  workflow.yml        # Step definitions
  research.xsd        # Schema for the "research" step
  research.md         # Prompt for the "research" step
  analysis.xsd        # Schema for the "analysis" step
  analysis.md         # Prompt for the "analysis" step

Each step needs a <step-id>.xsd (validation schema) and a <step-id>.md (prompt template). File names must match the step id.

Here is a minimal workflow.yml:

workflow: my-workflow
description: "Analyze a topic"
max_retries: 3

steps:
  - id: research
    description: "Gather key facts and sources"

  - id: analysis
    description: "Synthesize findings into recommendations"
    predecessor: research

Steps without a predecessor run first. Steps with a predecessor receive that step's output XML as context in their prompt.

Optional hooks:

• pre_script -- shell command to run before the LLM call (e.g., fetch data)
• post_script -- shell command to run after a valid output is written

CLI Reference

Command	Description
laisi init	Scaffold .laisi.yml and .laisi/ directory
laisi init --workflow <name>	Copy a built-in workflow to get started
laisi run	Execute the next pending step, then exit
laisi run --all	Run all remaining steps (stops on failure)
laisi run --step <id>	Run a specific step
laisi status	Show workflow progress

Requirements

• Node.js 20+
• claude CLI (Claude Code)

License

MIT