canon

canon

The canonical setup for Claude Code projects, with an experimental Codex port.

One command installs three pieces of community work and wires them together as one system: Andrej Karpathy's CLAUDE.md / MEMORY.md / ERRORS.md persistence trio, Every Inc's Compound Engineering /ce:plan + /ce:work planning loop, and Matt Van Horn's /last30days research skill. canon adds the bootstrap and the hooks that keep the discipline from rotting.

What it composes

Piece	Source	Role
CLAUDE.md / MEMORY.md / ERRORS.md trio	Karpathy	Behavioral spec + decision log + failure log. A small, human-written context file the agent reads at task start.
/ce:plan + /ce:work	Every Inc — Compound Engineering	Parallel research agents produce a structured plan; execution ticks off acceptance criteria.
/last30days	Matt Van Horn	Parallel community-knowledge search across Reddit / HN / Polymarket / GitHub / X / YouTube / TikTok / Instagram / Bluesky / open web — grounds /ce:plan in fresh source material. Most sources free; ScrapeCreators key unlocks paid platforms.

The canonical loop is research → plan → execute → persist:

/last30days <topic>     → research
/ce:plan <task>         → plan, grounded in fresh research
/ce:work                → execute
MEMORY.md auto-updates  → persist (via canon's Stop hook)

What the evidence says

canon's claim is deliberately narrow: a small, human-written context file helps a little; a large, machine-generated one hurts and costs more. That's not a slogan — it's the finding of the only rigorous study to date. ETH Zurich's Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents? (arXiv 2602.11988, 2026) tested four coding agents across 438 tasks under three conditions — no context file, an LLM-generated one, and a developer-written one. Developer-written files gave a modest real gain (+4% on their benchmark). LLM-generated files reduced task success in 5 of 8 settings (−3% vs no file at all) and pushed inference cost up by over 20%. Removing an "Architecture"/overview section, keeping only commands, constraints, and non-standard patterns, produced the same behavior at a lower token budget.

So canon optimizes for the version that works: keep context files small, human-authored, and pruned. The persistence trio is a human-curated log, not an auto-generated dump; the Stop hook proposes memory entries for you to confirm rather than writing them silently; and optimize can prune a context file and prove against an eval that the cut reduced cost without regressing behavior. canon's job is to keep you on the right side of that line — not to make a context file as big as possible.

What canon adds

The four pieces above are the prior art. canon is the glue: