Work in loops

Turn one large AI request into a sequence you can steer.

One giant request hides too many decisions.

The output arrives polished, but you cannot tell where the direction narrowed, which assumption entered the work, or what to change without starting again.

A loop keeps the work visible: set direction, make one pass, inspect it, then decide what the next pass deserves.

See it

Relay's designer asks AI to “analyse the research, write the brief, explore solutions, and produce the final handoff flow.”

What went in: Interview notes, a product request, and four jobs combined into one instruction.

What AI produced: A complete-looking flow built on an unsupported belief that managers should approve every handoff.

What the designer changed: They restarted as five passes: frame the problem, surface assumptions, explore directions, make one flow, check it against evidence. The approval assumption appeared in pass two and was removed before it reached the interface.

Why it works

  • Smaller passes make inputs and decisions inspectable.
  • A checkpoint is cheaper before an assumption reaches every downstream artefact.
  • Each output becomes useful material for the next pass.
  • The loop can tighten as confidence grows instead of repeating the whole task.

Try it

Try it · 6 min

Draw the loop

Choose one task that currently feels too large. Break it into three to five passes. For each pass, name the input, the output, what you will inspect, and whether you continue, revise, or stop.

KeepA loop map with a visible input, output, and human checkpoint for every pass.

Check it

  • Does each pass answer one question?
  • Can you inspect its output before the next pass begins?
  • Is source material added only where it helps?
  • Are the riskiest assumptions checked early?
  • Can the loop stop without wasting later work?
Human checkpoint

Where would a weak assumption become expensive if you let the loop continue?

Look for the pass that determines the user, problem, product model, or irreversible implementation direction. Put a human check immediately after it.

Keep it

WORK LOOP

PASS 1
- Question:
- Input:
- Output:
- Check:
- Continue when:

PASS 2
- Question:
- Input:
- Output:
- Check:
- Continue when:

FINAL DECISION I KEEP
-

Before you use this

More passes do not automatically make better work. Stop when another pass cannot change the decision. Protect source material at every handoff, and do not let AI mark its own unsupported assumption as verified.

Can you use it?

Question: What makes a sequence of prompts a loop?

Answer: Each pass produces something you inspect and use to decide whether and how the work continues.

Next

The first primitive inside a useful loop is context that tells AI what the work actually is.