Prompt Engineering

As models become more capable, prompt efficiency mostly comes down to communication skill and domain knowledge. The additional concern is sycophancy: language can steer a model into a destructive response. Counter is a repeatable prompt/skill intended to help guard against that failure mode.

The fundamentals are communication and domain knowledge

Accurate language lets you get more nuanced results from AI as a partner or team member. If you already know your domain and communicate clearly, there may be little conventional “prompt engineering” left to learn.

If communication is weak, the usual consequence is inefficiency: more turns to reach the desired output means more cost. If domain knowledge is weak, the risk is subtler. You may not be able to distinguish a passable response from a great one, so you cannot steer the model beyond “passable” or extract the full value of bespoke, highly customised software.

Skills package expert context

Expert knowledge is becoming more accessible through skills: fixed prompts with baked-in context. They can make otherwise scarce expertise more available at the moment of prompting.

Guard against sycophancy

Sycophancy is exaggerated, insincere flattery intended to gain advantage or approval. A useful shorthand is brown-nosing.

AI models can be highly sycophantic. Even a well-formed prompt can be undermined by a follow-up such as “are you sure?”: the model may retract an adequate answer because it defaults to being agreeable.

cold asks the model to base its answer on evidence and suppresses sycophantic language wholesale.
counter begins with a Big Five personality test. Its generation prompt uses those results to create a skill tailored to the user’s psychology and likely prompting failures.

The Counter repository includes an example skill, test results, and the research papers used as design directives.