How does this skill solve step skipping?

It uses a bash script to own the iteration loop, sending only one step at a time to Claude and verifying the result before moving to the next.

Can I resume a failed workflow?

Yes, the skill includes a --resume-from flag that allows you to pick up execution from a specific step index after a failure.

How are workflow steps verified?

The executor runs specific verification scripts that check the actual state of the filesystem, Git, or external APIs to confirm the work was performed.

What is the 'hallucinated completion' problem?

It occurs when an LLM claims to have finished a sequence of steps in a plan without actually executing the underlying commands or tools.

What is the difference between Standard and Serialized modes?

Standard mode sends each step individually for maximum reliability, while Serialized mode sends all steps at once to optimize for cost and latency.

Deterministic Workflow Executor

Name: Deterministic Workflow Executor
Author: fractary

$fractary$ byfractary

•

Productivity & Workflow

Orchestrates reliable, step-by-step AI workflows by shifting execution control from the LLM to deterministic bash scripts.

The Deterministic Workflow Executor (FABER) is a robust framework designed to eliminate the common 'hallucinated completion' problem where AI agents skip critical steps in a plan. By moving orchestration logic into a bash-based loop, this skill ensures Claude processes each task individually while validating real-world results through external verification scripts. It provides developers with guaranteed sequence adherence, precise state management using jq, and the flexibility to resume failed workflows from specific checkpoints, making it ideal for complex, multi-stage automation tasks.

Key Features

01Deterministic state management for precise tracking of workflow progress

02Support for both Standard (step-by-step) and Serialized execution modes

03Bash-driven orchestration ensures sequential execution without skipping steps

042 GitHub stars

05Real-world verification checks filesystem, Git, and GitHub state after each task

06Resume-from-step capability to recover from failures without restarting

Use Cases

01Automating multi-step codebase migrations that require strict sequential logic

02Managing long-running AI agent tasks prone to hallucinating completion

03Running complex CI/CD tasks where each stage must be verified before proceeding

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add fractary/claude-plugins deterministic-executor

For use in Claude.ai and ChatGPT

Download Skill