────────────────────────────────────────────────────────
THE LOOP GENERATOR
A prompt for inventing your own AI agent loops.
From "The Trustworthy-Loop Field Guide" by JAAX Labs
https://www.jaaxlabs.com/trustworthy-loops/
────────────────────────────────────────────────────────

WHY THIS EXISTS
A blank "give me ideas" prompt collapses to the same five answers every
time. A tight box of hard rules forces a model off the beaten path, and
running it through a model from a DIFFERENT family than you normally use
surfaces ideas your usual model's habits would never reach.

HOW TO USE IT  (5 steps)
1. Open a model from a different family than your daily driver.
   (If you live in Claude, run this in GPT, and vice-versa.)
2. Paste the prompt below.
3. Replace [YOUR SYSTEM] with a few honest lines about your real setup,
   your data, and what you're trying to achieve.
4. Run it 3-5 times. Each run, change the LENS line to a different angle.
5. Keep only the ideas that satisfy ALL FOUR rules. Bin the rest without mercy.

────────────────────────  COPY BELOW  ────────────────────────

Invent autonomous "loop" ideas for the system described below.

A loop is a system that prompts an AI agent on its own: it wakes on a
trigger, finds its own work, does it, checks its own result, and decides
whether to run again.

MY SYSTEM:
[YOUR SYSTEM, e.g. "A 2-person agency running paid ads for 3 clients.
Data lives in Google Ads + Shopify. Goal: grow accounts without me
watching dashboards all day."]

Generate 10 ideas. Every idea MUST satisfy all four rules, or discard it:

  1. NOVEL: not a known pattern. No "babysit a task until the tests pass,"
     no "cron job that sends an alert." If it already exists, it's dead.

  2. MACHINE-NATIVE: it must exploit something a human fundamentally can't:
     running hundreds in parallel, failing thousands of times cheaply,
     competing against itself, running 24/7. "A chore done faster" does NOT
     count.

  3. ROOTED: it must attach to MY actual system, data, or goals above.
     No generic ideas that would fit any company.

  4. GATE-NAMED: name the one thing that keeps each idea trustworthy:
     what stops it from quietly going wrong?

LENS FOR THIS RUN: [pick one and change it each run]
   - loops that exploit running many copies in parallel
   - loops that compete against or attack themselves to improve
   - loops that verify what is TRUE, not just what to do next
   - loops that exploit time you can't use (overnight, continuous, instant)
   - loops that make your knowledge base get better at being itself

Format each idea exactly like this:
   NAME: what it does (1-2 sentences)
   · exploits: <the machine-native trick>
   · gate: <the one thing that keeps it trustworthy>

────────────────────────  COPY ABOVE  ────────────────────────

ONE LAST RULE
Whatever it invents, don't trust the output until an independent check
(a different kind of model, or a hard test the agent can't edit) agrees.
The loop is the easy part. The verifier is the job.