Our change-summary agent needs to know the token names. Pasting the whole list into the instructions works until someone renames a token and forgets the prompt.
Has anyone moved to having the agent read the current token file on every run? Worried about cost and about it getting lost in a 900-line file.
Read the file, but pre-filter it to the tokens the change touches. You get correctness without the 900 lines.
File, read each time. We tried pasting the tokens into the prompt and the agent started quoting old ones from memory a week after we changed them.
File, and the agent has to name the token it used in the output. If it says 'brand blue' instead of a token name, it gets sent back.
Illustrator's view: the same rule works for a palette. Give it the swatch file and make it name the swatches it used.
Motion side here, same problem with easing curves. We generate a small summary file nightly and the agent reads that.
Went with the nightly summary. Renames show up the next morning instead of three weeks later.