Peer helps make a model behave more like a trusted colleague.
Its goal is to not give in at the first sign of pushback yet concede when the information warrants it.
It doesn't make it stubborn, just more willing to stand up to you and act as a sounding board, so you can talk things through without it defaulting to agreement.
It reduces flattery and waffle whilst focussing on thinking about the reply rather than shotgunning answers hoping one sticks.
I tested it blind with and without the skill. Over 144 pairs it gave in to pushback with no new argument once, against 15 times without it. In a second round, both sides got a decent set of custom instructions (216 pairs across Opus, Sonnet and Fable): Peer was judged better 132 times, the instructions alone 36, and 48 were ties.
The details of the testing are in the GitHub readme and the tests folder.
AndyKear•11m ago
Its goal is to not give in at the first sign of pushback yet concede when the information warrants it. It doesn't make it stubborn, just more willing to stand up to you and act as a sounding board, so you can talk things through without it defaulting to agreement.
It reduces flattery and waffle whilst focussing on thinking about the reply rather than shotgunning answers hoping one sticks.
I tested it blind with and without the skill. Over 144 pairs it gave in to pushback with no new argument once, against 15 times without it. In a second round, both sides got a decent set of custom instructions (216 pairs across Opus, Sonnet and Fable): Peer was judged better 132 times, the instructions alone 36, and 48 were ties.
The details of the testing are in the GitHub readme and the tests folder.