What is orchflows?

Orchflows is a set of composable skills for Claude Code and Codex that decides how much process a task actually needs. Instead of manually chaining separate commands or memorizing skill names, you just describe what you want in plain language, and the system routes the request to the smallest workflow that can prove it finished the job. A one line question gets a one line answer, while a large feature build gets a frozen specification, tickets broken into pieces, work split across parallel subagents, and a review step at the end. There are 37 small skills in total, and each one is kept intentionally short so it does not clutter the model's context with unnecessary instructions. The project is built around small reusable pieces called bricks, each doing one job. For example, orch-deliver ships work, orch-critique attacks it looking for flaws, orch-judge scores it without bias, orch-loop repeats a step until a condition passes, and orch-fix proves the cause of a bug before attempting a repair. These bricks connect through a small number of shared data formats, so a piece built for one kind of project, such as writing code, can be reused for another, such as a research report, just by swapping which pack of settings is active. A notable feature is self-improvement: a skill called orch-self-improve reads logs of past runs, including retries and workarounds, and turns recurring friction points into proposals the user can accept or reject, such as adding a missing documentation step to a workflow. There is also orch-visualize, which turns a workflow, codebase, or process into a Mermaid diagram that is checked against a real renderer before being returned. Installation is a git clone followed by a shell script that detects Python and configures whichever of Claude Code or Codex is present on the machine, and works on both Windows and POSIX systems. The README does not state a license in the portion available.