Two skills can each be perfect and the stack still fails because what one produces isn't what the next can consume. Most research stacks fail at the joints, not the nodes. Design them together.
The data-refiner says {bulk_sensitivity: 7.6, Q_loaded: 18200}
but the manuscript-weaver wants {prior_art_map, hypothesis, protocol, figure}.
So you're manually pasting fits into discussion sections every Wednesday. The fix is
designing the schemas together, not separately — or inserting a small connector skill
that translates the joint.
Take any two skills from the research production line. Write A's output schema and B's input schema side by side. If they don't match, either fix the schemas or write the connector skill. The joint IS the design — it's not an afterthought.