Plain words

Page 2 of 7

Every word this repo uses, explained once. If a page loses you, the reason is almost always on this list. The field guide has a longer version on page 6.

The five words that matter most

Loop. A small system that finds work, hands it to an agent, checks the result, writes down what it learned, and exits. You build the loop. The loop runs the agent.

Check. The script that answers one question: is there work? check.sh in every loop folder. Exit 0 means no, exit 1 means yes, and exit 2 means not wired: the check cannot see what it needs in this repo, so it fails loudly instead of guessing. Nothing is spent unless it says yes. That third answer matters more than it sounds: a check with only two will say "there is work" when it means "I cannot tell".

Orders. ORDERS.md in every loop folder. What this loop owns, what it must never touch, and when to stop and ask a person. The agent reads this every run.

Agent. A model that does not just answer. It reads your files, runs commands, edits code, checks its own work, and keeps going until the job is done. Claude Code or Codex, here.

Gate. The moment a human is, or is not, required. Ships on green means no human. Flags, you decide means a human owns the button.

The repo's own words

Chapter. One of the four folders under loop-packs/. Each holds nine or eight loops that share a runner and a set of rules. The guide calls them chapters; the folders are named after them.

Pack. The same thing as a chapter. The folder names end in -loops.

Standing orders. CLAUDE.md (or AGENTS.md for Codex) at the top of a chapter. The rules that apply to every loop in it, read on every single run.

Generated page. Every docs/*.html is rendered from the .md beside it. Edit the markdown; a push to main rebuilds the rest.

Memory. memory/NN-name.md. One line per run, written by the agent, kept in git so a wrong lesson is one revert away.

Demo app. demo-app/ inside a chapter. A small project deliberately seeded with problems, so you can watch the checks find real work before you point one at your own code.

loops.env. One file at your repo root holding the settings your loops read: the command that proves a bump is safe, the command that diffs your screenshots. ./install.sh writes it with every setting your chosen loops need, commented out with an example. Commit it; these are commands, not secrets.

run-loop.sh. Runs one loop on your machine, exactly as the workflow would: the check, then the agent only if there is work, then the check again. No secret and no schedule needed.

Dry run. DRY_RUN=1 in front of a cloud check. It runs against canned AWS responses in mock/, so you can try every cloud loop with no AWS account.

The general words

Repository (repo). The folder holding all the files for one project, plus its history.

Branch. A private copy of the project where changes can happen without touching the real thing.

Pull request (PR). A proposal: here are my changes, please review and merge them. The thing you approve.

CI. The robot that runs your tests automatically when changes are proposed. GitHub Actions, here.

Fence. The list of tools an agent is allowed to use, passed as --allowedTools. The orders are the soft rule. The fence is the one that is actually enforced.

Cap. --max-turns and --max-budget-usd. The run ends by itself before it can get expensive.

Token. The unit you pay for, roughly a word.

Context. What the agent holds in mind during one run. Fill it with junk and the agent gets slower, pricier, and worse.

← What a loop is · Contents · Next: Run the demos →

The 35 loops: Engineering · Cloud · AI and ML · QA