loops

Playbooks for AI coding agents in Cursor and Claude Code.

Say “use the loops” and the agent picks a workflow — plan, build, test, review — and follows it step by step. No extra app. No account. You install from this repo once.

Needs Node.js 18+.

Install (all projects on your computer)

git clone https://github.com/noidsoup/loops.git ~/Code/loops
cd ~/Code/loops
node adapters/emit.js
node adapters/install-global.js

Then open any project in Cursor or Claude Code and say use the loops, or just describe the job (“write tests for the parser”, “review this PR”).

In Claude Code you can also type /loops-dispatcher, /loops-tdd, and so on.

Update later:

cd ~/.loops && git pull && node adapters/emit.js && node adapters/install-global.js

Remove: node adapters/install-global.js --uninstall

What to say

You say	What happens
“use the loops”	Picks the best fit
“build a new auth flow”	Plan, then build
“write tests for the parser”	Tests first
“stress-test this design”	Spec, attack, fix gaps
“review this PR”	Pre-merge review
“reproduce this bug”	Repro → failing test → fix
“upgrade Next to 15”	Upgrade with a checklist
“explain this codebase”	Map of the repo
“remove the slop” / “de-ai-ify”	Clean up AI-sounding code
“swarm this”	Full start-to-finish pipeline

You usually don’t need the workflow name. Just describe what you want.

Which AI models to use

Out of the box, Cursor stays on Auto.

For the recommended stack (Opus on hard planning; lighter models elsewhere):

cp adapters/MODEL_CLASSES.local.example.md adapters/MODEL_CLASSES.local.md

That creates a gitignored MODEL_CLASSES.local.md on your machine. Details: adapters/MODEL_CLASSES.local.example.md. Soft defaults: adapters/MODEL_CLASSES.md.

One project only (optional)

If you want loops locked to a single repo for teammates:

cd your-project
git clone https://github.com/noidsoup/loops.git .loops
node .loops/adapters/emit.js
mkdir -p .cursor/rules .claude/rules
cp .loops/INSTALL.mdc .cursor/rules/loops.mdc
cp .loops/INSTALL-CLAUDE.md .claude/rules/loops.md

More detail: examples/install-in-fresh-project.md.

Workflows included

Name	For
plan-and-implement	New feature or refactor — plan, build, Judge, revise
tdd	Change driven by tests — Prove via command output
sar	Spec → attack → repair (simplest correct)
adversarial-gate	Review before merge (max 3 rounds)
reproduce-and-fix	Bug → test → fix → Prove green
migrate	Framework or version upgrade
explain-codebase	Learn an unfamiliar repo
de-ai-ify	Tone down AI-sounding code
swarm	Full pipeline
use-the-loop	Smallest mix that fits when it’s unclear

Self-correcting quality control

Every building loop works in three roles: a Builder writes, a Judge checks the work against real command output (not the Builder's own claims), and a Manager decides to revise or ship. If it's not right after 3 rounds, it stops and asks you instead of guessing.

The companion: my-robot

Loops is the methodology. my-robot installs the whole foundation it runs on — verify gate, memory files, semantic search, an Obsidian wiki, agent rules — into any repo in one script, and symlinks loops in automatically.

License

MIT · v0.1.4 · CHANGELOG.md