Cookbook¶
Quickstart: one /evolve round on a toy project¶
examples/quickstart/run.sh needs nothing but honed and Claude Code. It copies a tiny book-tracker project (a
CLAUDE.md with one wrong instruction, and one skill) into a new git repository, sets it up for /evolve, runs three
Claude Code sessions in it, then runs a round: reflections on each session, a curation, the proposal's diff, and the
report.
examples/quickstart/run.sh [FOLDER] # default: a new temporary folder
See examples/quickstart/README.md for what to expect.
Graphify pipelines¶
The example evolution report comes from cookbook/evolve.py: Graphify's skill and context evolved
over four rounds on the FastAPI repository. Each round ran coding tasks in Claude Code, Codex, Pi, and OpenCode,
reflected on every session, curated one candidate per harness, and merged them into a release. Its runs predate token
usage recording, so the Token usage tab is empty.
Runnable scripts in cookbook/, each a complete pipeline on the Graphify configuration in
tests/fastapi and the training sessions under logs/. They make real model calls.
| Script | What it does |
|---|---|
cookbook/live_reflect.py |
Imports the Graphify configuration (each harness's own hooks included) into a workspace, then reflects on sessions that ran Graphify with the Lesson schema and records the ratings. |
cookbook/live_ace.py |
One ACE round: the same reflections, then one curation over all of them, ending with the configuration diff. |
cookbook/live_ace_with_merge.py |
Parallel pipelines: each in its own worktree, reflecting on its own sessions and curating; a merger combines the candidates and the run ends with the diff from the base. |
.venv/bin/python cookbook/live_ace_with_merge.py
Results go under artifacts/live-reflection/: the workspace (graphify/), the session index,
saved reflections, and — for merged runs — one directory per run with its worktrees.