Inspiration
When you ask an AI to do the same thing to a lot of items (reply to 50 emails, fix 40 files, rename 300 photos), you either check every result yourself or trust it and hope it understood you. If it didn't, every item comes out wrong the same way. Most people already say "do 3, show me, then do the rest." /canary turns that into a command, like a canary sent into a coal mine first.
What it does
Run /canary 3 "draft replies to all of these". The agent does 3 items that differ from each other and stops. You approve them or fix them, and every fix is saved as a one-line rule. Then it does the rest in growing batches, checking each item against your approved samples and rules. Anything that doesn't match gets held with a short reason instead of guessed at.
In the demo, it drafts 10 club email replies and holds 2: a formal note from the principal (the samples were casual) and a reimbursement question (the samples never touched money). It works on code too: it shows 3 routes as diffs on a new git branch, learns "use logError(), not console.error", applies it to the other 21 routes, and holds a legacy file that uses callbacks.
How I built it
/canary is one folder in the Agent Skills format, so it runs in Claude Code (installable as a plugin), Devin, Codex, and any agent that can read a markdown file. SKILL.md holds the instructions the agent follows. scripts/canary.py, standard-library Python 3.9+, keeps the ledger: approved samples, rules, held items with reasons, drafts, and a log of every decision in .canary/. The agent makes every judgment call and the script only does bookkeeping, which keeps it agent-agnostic. Samples are done in an undoable way, like drafts instead of sent emails. The ledger script has unit tests.
How it differs from just giving examples
Examples help, but the AI still runs everything in one go and fills gaps with guesses. /canary stops for review, writes your edits down as rules you can read and change, runs in batches, and holds the items it isn't sure about.
GitHub: stffinfcti
Discord: waxhy
Name: Ricardo Medrano


Log in or sign up for Devpost to join the conversation.