The PTMRO loop

Build-along · ~10 min at your own keyboard

PTMRO is a diagnostic lens, not a law: five things to look for when a run goes wrong — did it plan, which tools did it call, what did it keep track of, did it check its results, how did it decide the next step. Not every run shows all five, and a run that narrates a plan isn't better than one that quietly gets it right. Judge by what you can observe — tool calls, files changed, checks performed, and whether the result is correct — and use the lens to find WHERE a bad run went wrong.

The long version: How agents work ↗

Do this

0/4

Checkpoints

▸You should now see tool calls appear in the run, such as file reads or a script.
▸You should now see your check of the result against the source.
▸You should now see a note from you naming where a problem started, using the lens.
▸You should now see a visible change in the re-run after your one added instruction.

The three ways this goes wrong

Claude finishes in one step with no plan shown, and the result is right. — That's fine. The lens is for finding problems, not for requiring a narrated plan. Try a harder task if you want practice.
You cannot tell whether it checked its work. — Look for any point where it re-read its output or compared a number to the source. If there is none and the result has errors, that's where to add an instruction.
The re-run looks the same as the first. — Your added instruction was too vague. Make it a checkable action, like 'compare totals to the source file before you say done'.

Share what happened

Paste your five-beat observation notes and what changed after your one instruction.

Strip API keys, passwords, and tokens before you paste. Swap them for [KEY].

🔒 Don't paste confidential, personal-health, or payment info.