Vocabulary for writing skills and other agent-facing documents, drawn from mattpocock/skills and cursor/plugins pstack.
A
A/B eval
Running two versions of a document through the same task with the same model and tools, differing only in the edited text, to see which behaves better.
C
Context pointer
A reference held in the agent's context that names out-of-context material and encodes when to reach it. A skill's description is one. Its wording, not its target, decides how reliably the agent follows it.
Confound
A second change bundled with the one you meant to test, so a result cannot be credited to the word alone.
L
Leading word
A compact concept already in the model's pretraining (tight, red) that the agent thinks with while running a document. Repeated as a token, never as a sentence. Works in the body for execution and in a pointer for invocation.
N
Negation
Steering by prohibition. It puts the forbidden behaviour into context and makes it more available. Prompt the positive target instead.
No-op
A line that changes no behaviour because the model already does it by default. Test: delete it and see whether behaviour changes. Delete the whole sentence, not a few words.
Near-miss prompt
A prompt that belongs to a neighbouring skill. An eval uses it with not-skill-used to test a trigger boundary.
P
Predictability
The agent taking the same process every run, not producing the same output. The root virtue every lever serves.
Proxy
A measurable stand-in for the behaviour you care about, for example Bash before Edit standing in for a loop built before the fix. It can pass without the real behaviour.
S
skill-used
A promptfoo assertion that passes when the agent actually invoked the named skill.