Prompt or loop? — the shift
Skills you'll gain
Tell a prompt from a loop · apply the three-question test · write a verifiable stop criterion · pick among the four loop types · set up your first routine without taking on risk.
Key concept — the prompt describes, the loop verifies
With a prompt, you are the judge: Claude produces, you read, you correct, you run it again. You are the loop — and your attention is the bottleneck. With a loop, the machine is the judge: you define a verifiable target, Claude produces, tests, retries, and stops only once the target is met.
The line that captures the shift comes from Boris Cherny, head of Claude Code at Anthropic: "I don't prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops."
The three-question test
Before asking yourself "how do I prompt this?", run these three questions in order. The first one settles it nine times out of ten.
When to stick with a classic prompt
A loop isn't a promotion — it's a different tool. For most of your day-to-day work, the prompt is still the right reflex.
💬 Thinking, weighing, deciding
There is no automated test for "good decision." Your judgment is what settles it — and that's exactly what you're there for.
✍️ Writing something sensitive
A delicate email, a keynote, a message to an unhappy client. The success criterion is how the reader feels: that doesn't compile.
🔍 Exploring a new topic
When you don't yet know what you're looking for, a conversation beats a poorly aimed loop spinning in the dark.
⚡ A one-off, short task
Summarize a document, rewrite a paragraph, pull out figures. Building a loop would cost more than doing the task.
When to switch to a loop
🔁 The task comes back every week
Competitive monitoring, pipeline review, reporting, screening applications. You redo it identically: automate the trigger.
🎯 The result is measurable
A score, a threshold, a format to respect, a list to complete. The moment a number or a rule decides, the machine can judge alone.
⏳ The project is long
Data migration, auditing a document archive, compliance work. Opus 5 and Fable 5.1 sustains multi-hour tasks without losing the thread.
🧹 The work is repetitive and high-volume
200 product sheets to normalize, 80 contracts to index. The cost of one more iteration becomes negligible: let it run.
The four loop types
Anthropic distinguishes four. The first two fire on your request; the last two run without you.
| Type | Fires when… | Stops when… | What it's for |
|---|---|---|---|
| Turn-based | you ask | Claude judges the task complete | The default mode — the one you already use without naming it. |
| Goal-based | you ask | the criterion is met, or the turn cap is hit | "Keep going until this is true." The most useful one day to day. |
| Time-based | a schedule or interval | you cancel, or the work is done | Monitoring, reporting, syncing. Cowork's scheduled tasks. |
| Proactive | an outside event | each pass succeeds | An email lands, a file changes, a threshold is crossed: it fires itself. |
Putting it into practice — in five steps
Spot the task that comes back. Look at the week behind you: what did you do twice or more, identically? That's your first candidate. Don't invent a loop for a task you never actually do.
Write the stop criterion before the prompt. This is the single most important mental switch in this module. One sentence, verifiable, with no adjectives: "every entry has a price, a date, and a clickable source."
Have it stop at a queue. The loop prepares, sorts, proposes — you review and you publish. Never the other way around, at least until you have several weeks of experience with it.
Let it run for a week, then read it back. A loop is tuned in use: you'll find the criterion was too loose, or too strict. That's normal — and that's where the value gets made.
Watch the spend. A running loop consumes continuously. In Claude Code, the /usage command breaks down consumption by skill, by subagent and by loop — you see exactly what each loop costs.
2 · The loop that acts in the real world. Publishing, sending, paying, deleting: none of it can be walked back. A loop that prepares is a major win; a loop that publishes is a risk nothing justifies at the start. Have it stop right before the button.
After: a weekly loop produces three drafts — the summary note, the list of payment reminders, the agenda — and stops there. Stop criterion: "all three files exist, every figure is sourced, no reminder concerns an invoice already paid." You show up Monday, you read, you fix two lines, you send. 40 minutes become 8.
List three tasks you did twice last week. For each, try to write the stop criterion without an adjective.
Expected result : At least one task has a verifiable criterion: that is your first loop candidate.
Going further
This module answers "prompt or loop?". To build the loop once you've decided, move on to module 15 bis — the 5 patterns that work (chaining, routing, evaluator-optimizer) and module 15 — building AI agents. Scheduled tasks live in Cowork (module 05), and the /loop, /goal and /schedule commands live in Claude Code (module 06). Official reference: Loop engineering — the Anthropic blog.
→ Your best first loop is one where you already know the expected result: you'll see immediately if it drifts.
→ Ask Claude: "Turn this recurring task into a loop: propose the trigger, the stop criterion, and the guardrail."
→ A loop that never stops is a badly written loop, not an ambitious one.