* When I provide long, detailed protocols in prompts or CLAUDE.md / AGENTS.md files (e.g. make a plan, implement, test, debug, git commit, etc...) the agent will often skip or handwave steps.
* Often I'll find myself repeating the same patterns of prompts repeatedly. Examples: "diagnose the error" followed by "fix it", looping back and forth between "implement this spec" and "audit the implementation against the spec"
* The agents are fairly limited in terms of scope and max time spent on a per-prompt basis. This makes it challenging to set up long autonomous runs, e.g. overnight.
Today I'm happy to share *FlowCoder*, the project I've been working on to address these issues. FlowCoder allows you to create and execute custom automated workflows for Claude Code and Codex, via a visual flowchart builder. I am hoping this project can both help vibe coders scale their results and enable autonomous agent research by building on top of existing coding agents.
FlowCoder lets you set up slash commands to execute flowcharts of prompts and bash commands. These flowcharts have a fair number of features:
* The core building blocks are Prompt blocks, which send prompts to Claude Code or Codex, and Bash blocks, which run bash commands.
* FlowCoder keeps track of variables while executing flowcharts. Prompt blocks allow you to enforce the agent to respond with structured output to assign variables values, and Bash blocks allow you to save the bash output and/or exit code to variables.
* Branch blocks let you configure a boolean expression with these variables, splitting the flowchart into True and False paths.
* Flowcharts can accept CLI-style string arguments, and all blocks support syntax for argument substituion and variable substitution. So for example, you can create a prompt block that says "Create a spec for this task: $1" and it will substitute the first argument you pass in. README explains more.
* Command blocks allow you to call other slash commands from within your flowchart. FlowCoder maintains a stack of flowcharts to handle command recursion.
* Flowcharts also support Refresh blocks for resetting context and Variable blocks for initializing/setting variables.
* FlowCoder automatically creates a git commit after each Prompt or Bash block.
I've included a number of examples in the repo to help users get acquainted with the system, showcasing prompting paradigms like implement-audit loops and test-fixing loops, and programmatic paradigms like for-loop behavior. README explains more.
Note that these example flowcharts are not "optimized". They are a starting point. Flowcharts provide a huge amount of expressive power. You can encode the specifics of how you like to approach your software engineering practice, whether you prefer to vibe code in small chunks or large autonomous sequences. I have my own set of flowcharts I've been developing for my own practice, and I've seen significant gains as I've been going through the process of optimizing these flowcharts' structures and prompts.
I hope others can benefit from this work or may want to contribute! The project is still very young (v0). The codebase is in alpha and should be assumed to be UNSTABLE. It has been tested on Linux and WSL. Feel free to post any issues you encounter on the GitHub. Currently, I am using this version of FlowCoder to develop the next version of FlowCoder, an Electron-based version with a better-planned architecture and additional features (multi-agent/parallel workflows, CLI, UI improvements).
More info: https://youtu.be/1COOR6UmpsY