Reference card
The named smells, and the words for them.
Everything grading can catch, readable in full rather than only met on the way to a low score. Reachable from anywhere.
Role theater
Superlatives about the role add no constraint the model can act on.
State the job and its standard: what it produces and what makes the output correct.
AI self-reference
Spends tokens telling the model something it already knows.
Delete it. Start with the task.
Politeness padding
Courtesy tokens carry no instruction and dilute the ones that do.
Use the imperative: 'Return a JSON object', not 'Please return a JSON object'.
Emphasis inflation
When everything is critical, nothing is. Emphasis competes with itself.
Rank instead: put the binding constraint first, or state the consequence of breaking it.
Shouting
Capitalisation is not a priority channel; it just changes tokenisation.
Use a heading or put the rule in its own section.
Unbounded effort
Directly fights the budget: it asks for maximum output regardless of need.
Bound it. Say how long, how many items, or what counts as done.
Meta commentary
Framing around an instruction, rather than the instruction.
Drop the frame and keep the rule.
Catch-all ending
Signals a list the writer abandoned; the model cannot infer the missing members.
Either finish the list or replace it with the rule that generates it.