Prompt Engineering at Work
Move from one-off prompts to repeatable systems — templates you reuse, reliable ways to summarise long documents, a method for fixing prompts that go wrong, and a shared library your whole team can lean on.
Debugging a prompt that misbehaves
Debug like an engineer: change one thing at a time When a prompt gives you the wrong thing, the temptation is to rewrite it wholesale. Resist that. The fast route to a working prompt is a tight diagnostic loop, and its golden rule is: change only one thing between tries. Read what the model actually […]
Writing for a program, not a person
Nobody is watching when this one runs A prompt you run in a chat window has a safety net: you. You see a strange answer and retry. The moment a prompt runs inside a script, a scheduled job or a product feature, that net is gone — the output goes straight to a database, a […]
Handling failure on purpose
The call will fail. Decide now what happens then Model calls fail in ways ordinary functions do not: they time out, they get rate-limited, they return the right shape with the wrong content, and occasionally they return something no schema anticipated. A pipeline that assumes success is a pipeline that will corrupt data quietly. Four […]
A test set beats an opinion
“That feels better” is not a result Prompt work goes wrong in a specific, recognisable way: someone tweaks the wording, tries two inputs, decides it improved, and ships. Two weeks later the outputs are worse on a case nobody thought to try, and there is no way to tell which of the eleven edits since […]