Doctrine beats persona — what 259 production system prompts teach
evidence-graded · adversarially-verified
A synthesis of 259 shipping system prompts (Claude, GPT-5.x, Gemini, Cursor, Grok and more), anchored to a blind A/B/C/D test and distilled into seven cross-vendor convergences and seven optimization laws. The lever is rules + anti-scope + an output contract — not a richer character.
A 278-word doctrine (operational rules + one anti-scope line + an output contract) scored 1.000 — beating a 250-word pure persona at 0.944 and a 34-word bare prompt at 0.889. At equal length, doctrine beat persona; brevity lost to both.
Independent vendors converge: write identity in the third person by name so it survives role-play; default to read-only and act only on an explicit instruction; make the final message the only delivery channel; treat external/retrieved text as data, not instructions.
Verify the real surface, scaled to blast radius — CI-green is necessary, not sufficient (“running tests proves you can run CI, not that the change works”).
Persona is not binary: kill motivational pep-talk, but ship a product voice through one sealed slot with a show-don’t-tell seal, carved out of user-shipped artifacts.
Durable memory must store declarative facts, never imperatives that re-fire later as standing orders overriding the live request.