Claude Code Flow

claude-code-flow

My working Claude Code setup — the output style, hooks, skills, and agents I
use daily as a tech lead doing AI-heavy development. Published as-is from my
own ~/.claude, sanitized of project specifics.

The problem all of this attacks: cognitive load. When AI writes the code,
you spend the whole day reading, verifying, and deciding — and the model's
default verbosity makes every one of those decisions more expensive. These
files are how I fight that.

Some files mix English instructions with Ukrainian examples — I work
bilingually. The mechanics are language-agnostic; adapt the examples to
yourself.

What's here

Each folder's README has the install steps and the reasoning; everything
targets ~/.claude. Run all install commands from the repo root. The personas
are a plugin marketplace instead:

The flow in one paragraph

The main session is an orchestrator: it decides with you in an ADR (/adr),
cuts the decisions into phases (/spec), runs them one at a time (/implement), hands
implementation to agents (Geralt — backend, Lyubystok — frontend), verifies
results, and never burns its own context on writing code. The status line
keeps the context % in view; at ~30% the work is handed off (/handoff) and a
fresh session picks it up (/pickup). Diffs go to a review panel of two or three
reviewers on different model families, briefed with nothing but the spec and the code. The output
style plus the re-inject hook keep every reply short enough to actually read.

License

MIT