Loading the gallery
Commands that review a prompt, a tool definition or an agent brief before you send it, and Frameworks that hold the practices they check against, each drawn from a cited source.
Commands that run a check you would otherwise do by hand
Review a prompt before you send it
Check a prompt you are about to send against the rules for writing one a model follows
Review a prompt bound for a reasoning model
Check a prompt against the rules for reasoning models, where chat-model habits work against you
Review a tool definition an agent will call
Check a tool definition for the things that make an agent call it wrongly or not at all
Each one gathers the rules a job needs, from cited sources
Assume everything in the prompt will be seen
Write a hidden prompt on the assumption that a user will eventually read every word of it
Brief an agent so it does not guess
Give an agent enough context to act without inventing the parts you left out
Choose examples that do not skew the answer
Pick and order few-shot examples so the model learns the task rather than the distribution
Define a tool an agent calls correctly
Describe a tool so a model calls the right one, with the right arguments, and recovers when it does not
Give the model the data it needs
Put current, well-shaped data into a prompt without spending the whole context on it
How to instruct a model
Write the instruction half of a prompt so a model follows it
Pick the model the task actually needs
Match a task to a model class rather than reaching for the largest one by default
Prompt a reasoning model
Instruct a model that reasons before it answers, where the habits that help a chat model hurt
Write an instruction a model follows the first time
Turn a rough request into an instruction a model can act on without a second attempt
Every rule in a Framework names the source it was drawn from, and the card links to each source. Check a rule against the text before you rely on it.
Save a card into your own workspace and adapt it to how your team works. The license each source carries is shown on the card.
A Command is a prompt that runs one check, such as reviewing a prompt before you send it. Its card lists what it checks for.