Skills, references, and command-line tools for working on the Flow programming language with an AI coding assistant.
Flow moves quickly and it is not in any model's training data. Without help, an assistant will invent syntax, pick the wrong compiler host, and report success on a program that never ran. This pack gives it the current facts and a working method.
Install it once, then ask for what you want in ordinary language.
git clone https://github.com/flooooooooooow/flow-skills.git
cd flow-skills
./install.shThe installer:
- copies every skill into
~/.claude/skills/, - links
flow-doctor,flow-find,flow-check, andflow-newinto~/.local/bin, - prints what it did and what it skipped.
Pass --link to symlink the skills instead of copying, which is what you want
if you plan to edit them. Pass --dry-run to see the plan without touching
anything.
Uninstall with ./install.sh --uninstall.
- Flow, installed with
brew install flooooooooooow/flow/flowor cloned from source. - A C compiler.
xcode-select --installon macOS,build-essentialon Debian. - Python 3.9 or newer, for the full-language compiler host.
Run flow-doctor at any point to check all of this.
| Skill | Use it for |
|---|---|
flow-setup |
Installing Flow, checking the toolchain, fixing a broken environment |
flow-basics |
Writing and running an ordinary Flow program |
flow-dynamics |
flow blocks, evolves as, events, simulations, control |
flow-effects |
Algebraic effects, handlers, dependency injection, testing seams |
flow-autodiff |
Gradients, optimisation, small neural networks |
flow-graphics |
Windows, drawing, animation, recording GIFs |
flow-audio |
Synthesis, DSP, real-time audio safety |
flow-wasm |
Compiling a program to run in a browser |
flow-stdlib |
Finding existing library code before writing new code |
flow-debug |
Diagnosing a failure at the stage where it actually happened |
flow-verify |
Testing, evidence, proving a program does what it claims |
flow-perf |
Benchmarking and making a program faster |
flow-compiler-dev |
Changing the compiler, runtime, or standard library |
Each skill states which compiler host it needs, gives working code, and ends with the verification step that proves the work.
| Command | What it does |
|---|---|
flow-doctor |
Checks toolchain, hosts, and runs a hello-world smoke test |
flow-find <name> |
Searches the standard library and examples for a symbol |
flow-check <file> |
Compiles and runs a file, choosing the host automatically, and prints the exact command that worked |
flow-new <kind> <file> |
Scaffolds a program from a template: basic, dynamics, effects, gfx |
They all resolve the Flow checkout through $FLOW_HOME, falling back to
~/flow and then to whatever flow is on your path.
Flow has two compiler hosts.
| Host | Command | Covers |
|---|---|---|
flowc (default) |
flow run f.flow |
Functions, structs, arrays, pointers, loops, match, C interop |
python |
FLOW_HOST=python flow run f.flow |
Everything, including println, effects, flow blocks, generics, MLIR, GPU |
If you see flowc emit failed (Stage-A subset?), the program needs the Python
host. Most people should set export FLOW_HOST=python in their shell and stop
thinking about it.
AGENTS.md at the root of this repo is the portable version of the rules. Copy
it into a project folder and point Cursor, Codex, Copilot, or anything else at
it. It is plain Markdown with no tool-specific syntax.
references/ holds the same material as flat documents:
syntax-card.mdis the whole language on a few pages.command-map.mdmaps a goal to a command.stdlib-index.mdlists every standard library module and what it is for.error-map.mdmaps an error message to its cause.
The repository is the truth, and this pack is a map of it. When the two
disagree, the repository wins. Every skill tells the assistant to check the
current behaviour with flow help, a grep, or a test run before relying on
what is written here.
The deep background is Working with AI on Flow, the operating handbook in the main repository. This pack is that handbook cut into pieces an assistant can load on demand.
MIT. See LICENSE.