To reach agreement, international negotiators often compromise by using flexible language: they make controversial provisions vague, or add options and caveats. Does flexibility in agreement language influence subsequent state behavior? If so, do states follow both firm and flexible language somewhat, as negotiators hope? Or do governments respond strategically, increasing their energies on firmly specified tasks, and reducing their efforts on flexibly specified ones? Testing theories about agreement language is difficult because states often reserve flexible language for controversial provisions. To make causal claims, we study an unusually drafted agreement in which states had almost no opportunity to dilute agreement language. We examine the influence of the 1991 Paris Principles on the Design of National Human Rights Institutions (NHRIs), using an original data set of twenty-two institutional safeguards of NHRIs in 107 countries, and case studies. We find that variations in agreement language can have large effects on state behavior, even when the entire agreement is nonbinding. Both democracies and authoritarian states followed the principles' firm terms closely. However, authoritarian states either ignored or reduced their efforts on flexibly specified tasks. If flexibly specifying a task is no different from omitting it altogether, as our data suggest, the costs of compromise are much greater than previously believed.