Understanding Loops CHAPTER 03 Β· THE CALL
The Decision

Do you need a loop?

Honest answer: often, no. Loops are powerful, and like most powerful tools they're frequently unnecessary. Four boxes decide whether a task deserves one β€” and when it doesn't, there are two other routes that fit better. Ten seconds of checking replaces an hour of ceremony.

The four-box test

πŸ”
It repeats
you'll do this again β€” weekly report, every product, each client
🚫
Bad output is catchable
something can automatically reject a wrong draft β€” a count, a match, a test
πŸ€–
AI can run it end-to-end
no step in the middle needs your hands or your approval to continue
🎯
"Done" is objective
two people would agree it's finished β€” no taste call required
βœ“ ALL FOUR β†’ LOOP IT. MISSING ONE β†’ ONE OF THE TWO ROUTES BELOW.
The four-box test: does this task deserve a loop?

Box two is the gatekeeper. A loop is only as good as its checks, and checks must be falsifiable β€” countable, matchable, verifiable. "Under 120 words" qualifies; "sounds like me" doesn't, and no loop can save it. Chapter 02 showed what happens when checks go soft: accuracy fell off a cliff and correct work got destroyed.

There's also a shortcut version of this test, courtesy of the practitioners: the moment you catch yourself thinking "I'll have to try a lot of variations," that's the loop signal. Your instinct has already answered the four boxes.

πŸ”“
The unlockThird re-prompt = the test just ran itself. If you're fixing the same task for the third time, it repeats, it's catchable, and it's loopable β€” you're doing the machine's job.

Three routes, honestly assigned

ROUTE 1 Β· FULL LOOP β€” passes all four boxes

Work with hard requirements someone will hold you to: the description that must name the price, the report that must cite every figure, the code that must pass tests, the Friday deliverable. Write the six-label file from Chapter 02 and let it run. If it repeats, the five-minute setup amortizes to zero.

ROUTE 2 Β· CANDIDATES & PICK β€” short artifacts

Names, headlines, hooks, subject lines, one-liners. Don't refine one draft in circles β€” generate three independent candidates, grade them against your criteria, keep the winner. For short work this measurably beats polishing the same draft, and it takes one message:

Give me 3 independent options for [the thing].
Then grade each against: [your 2–3 criteria].
Tell me which wins and why, quoting the evidence.
ROUTE 3 Β· ITERATE LIVE β€” taste work

Design, voice, creative direction β€” anywhere "done" is a judgment call. Rubric theater doesn't help here; react to what you see, one round at a time, like an editor. The craftsmanship layer (length, structure, required beats) can still be looped β€” the taste calls stay yours.

ROUTE 0 Β· JUST A SPEC β€” most everyday asks

Quick answers, brainstorms, casual emails, summaries you'll skim. The four decisions from Chapter 01 buy a strong single attempt β€” take it and move on. Adding a loop here is bureaucracy.

The routes: match the tool to the task, not the hype
πŸ”“
The unlockCandidates-and-pick is the most under-used move in AI. For anything short, three graded options beats ten rounds of "make it better" β€” and nobody taught you because it isn't glamorous.

The borderline cases

Two situations genuinely straddle the routes. Creative work: you can't check "is this moving?" β€” but you can check structure: length, point of view, required beats, banned clichΓ©s. Run a loop for the checkable layer, then judge the result like an editor. Research and analysis: "is this true?" is hard to self-check, but "does every claim have a source?" and "are all five competitors covered?" are countable. Loop the coverage and the sourcing; keep the judgment.

πŸ”“
The unlockSplit the layers. Loop what can be counted; keep what must be felt. That one sentence resolves nearly every "can I loop this?" dilemma.

A fair warning about overuse

Once loops click, the temptation is to loop everything β€” and an unnecessary loop isn't just wasted setup. It adds passes where one attempt was fine, manufactures checks for things that didn't need checking, and slowly teaches you to distrust good-enough first drafts. The builders quoted in Chapter 02 don't loop their lunch orders. The skill isn't running loops; it's knowing which work deserves one.