Analysis updated 2026-05-18
Turn a written SOP document into a repeatable, code-driven workflow instead of ad hoc AI prompting.
Catch missing information with placeholder tokens instead of letting the AI invent facts.
Require human approval before treating any AI-produced output as final.
Audit an existing workflow for complexity and correctness without changing it.
| s0912758806p/agentic-sop-to-work | opennswm-lab/faros | pydantic/httpx2 | |
|---|---|---|---|
| Stars | 173 | 174 | 176 |
| Language | Python | Python | Python |
| Setup difficulty | easy | hard | easy |
| Complexity | 3/5 | 4/5 | 2/5 |
| Audience | ops devops | researcher | developer |
Figures from each repo's GitHub metadata at analysis time.
Needs python3 3.8+ available, installs via three commands in the Claude Code plugin marketplace.
agentic-sop-to-work is a Claude Code plugin that takes a written process document (a standard operating procedure, or SOP) and converts it into an automated workflow that an AI agent can run reliably and repeatedly. The project is aimed at work where mistakes are expensive: regulated processes, compliance tasks, anything where the output needs to be verifiable rather than just plausible. The core concern the plugin addresses is that AI agents left to their own devices tend to invent details when information is missing, make decisions autonomously when they should pause for a human, and gradually accumulate complexity until nobody knows what they are doing. The plugin addresses each of these by design. Any fact not present in the input is marked with a placeholder token rather than guessed. Control flow decisions are written in code rather than delegated to the language model. Every output is labeled as a draft that requires human approval before being treated as final. A Stop hook runs a regression check every time a skill changes to prevent the workflow from drifting. The engine behind the plugin is a Python script called run.py that reads a workflow definition in JSON format and executes steps one at a time. Each step passes through a set of gates before the next one runs: a command gate, a schema gate, a trace gate that checks no information was invented, and a recompute gate. If a gate fails, the system stops and asks for human input rather than patching the output to force a pass. A dry-run mode lists every planned step and validates any branching logic before anything is executed. Installation in Claude Code takes three commands in the chat interface to add the plugin from the marketplace, install it, and reload plugins. Python 3.8 or later must be available as python3. The kit directory can also be copied into any project directly, with a bootstrap script to scaffold new steps and a tool to export them as skills that the Claude Code agent can trigger by detecting the user's intent in conversation. The plugin ships with two skills: one that accepts either a plain description of a need or an existing spec and produces a workflow from it, and one that audits an existing workflow for complexity and correctness without modifying anything.
A Claude Code plugin that converts a written standard operating procedure into a gated, repeatable automated workflow for AI agents, aimed at regulated or compliance-sensitive work.
Mainly Python. The stack also includes Python, Claude Code, JSON.
Setup difficulty is rated easy, with roughly 30min to a first successful run.
Mainly ops devops.
This repo across BitVibe Labs
Verify against the repo before relying on details.