@curious_vii Do you think wiring the transcript up to a fast model like composer or 5.5 fast could do a reasonable job on the fly or does it need more steering?
when evaluating new products i start thinking in tokens
if every $20/mo subscription represents somewhere between 100m and a billion tokens, and the off the shelf software isnt getting the job done, then yea people are going to start building their own internal solutions
core primitives for quality ai work today:
-STT (inline & calls)
-AI questioning you
-managing & mining context
-clear success criteria
-evals & continuous iteration
-applying compute
@TaylorPearsonMe “it's just a bad catch all that is going to clog up context and get lost”
If you’re good at steering models then that’s basically what happens imo
@SteveWiesnerSMB skills are more for guidance & custom instructions and MCP is built for pulling in context (data) from external application
skills have the added benefit of only being fully called when needed
@curious_vii Maybe phrased differently:
It was easier to codify knowledge into techne.
Bc techne was high leverage (and still is) and the knowledge of how to turn something into techne was expensive and scarce