Orchflows
A composable skill library that turns Claude Code or Codex into an orchestrator with parallel subagents, verification, and self-improving workflows.
DanMcInerney · AI review on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add DanMcInerney/orchflowsAbout this skill
orchflows
A skill library that turns Claude Code or Codex into an orchestrator: subagents that work in parallel, real verification of everything they produce, and workflows that improve themselves. You just talk; it picks the right workflow.
Install
git clone https://github.com/DanMcInerney/orchflows
cd orchflows
./install.sh # install.cmd on Windows
Works with Claude Code and Codex, on Windows, macOS, and Linux. It
configures whichever CLI it finds. To update: git pull, rerun.
Use
Ask for things like you normally would. There are no skill names to memorize.
> why did memory double last release?
# a cited answer from a read-only research lane
> ship dark mode
# a frozen spec, tickets, parallel subagents, one review gate,
# and a final verification against what you asked for
> fix the flaky login test
# reproduce it, prove the cause, repair it, guard the repair
> research the top three auth libraries, then integrate the winner
# a research run chained into a code run, automatically
Small requests stay small — a one-line question costs a one-line answer. Big requests get real project structure. Nothing marks "done" except an external check passing.
More ways to use it
Routing has four branches — answer, ad-hoc, deliver, fix — and picks
the smallest one that can prove it's done. Everything else runs only
when you name it, so the routing table never grows as the library
does. The routing law is rules/topology.md §2, installed at
~/.orchflows/host-block.md. If routing gets in the way, orch-off
stands it down for the session.
Name the bricks yourself when you want a specific shape:
> orch-loop orch-deliver until `pytest -q` exits 0
> orch-panel these three cache designs — blind judges, pick one
> evolve the summarizer prompt against the frozen benchmark
Or build your own workflow in plain English:
> build me a workflow that does spec