Stop telling your model who it is
Every prompt-engineering guide says "assign a role." We tested it. Personified prompts — "you are a conceptual reviewer" — underperform plain task descriptions on both recall and cost. The leaner the persona, the better the model.
The orthodoxy
"You are an expert software engineer." "Act as a meticulous code reviewer." The personified system prompt is everywhere — in frameworks, in guides, in most production agent stacks. The theory: assigning a persona focuses the model on the relevant expertise and suppresses irrelevant outputs.
We assumed this was right. Then we tested it.
The probe
Same role (conceptual reviewer), same items (planted conceptual flaws), same five models, same MCP context. The only variable: prompt style. Three variants:
- PERSONIFIED — "You are a conceptual reviewer. Your job is to..."
- GENERAL — "Task: review the following for conceptual soundness. Method: enumerate assumptions, stress each..." (depersonalized, method described)
- ULTRA-LEAN — "Find conceptual flaws. Output: {assumption, verdict, fix}." (one-line goal + output contract)
The results
| Variant | Recall | Precision | Avg prompt tokens | Avg reasoning tokens |
|---|---|---|---|---|
| PERSONIFIED | 53% (8/15) | 62% | 548 | 1,433 |
| GENERAL | 73% (11/15) | 79% | 362 | 1,819 |
| ULTRA-LEAN | 87% (13/15) | 65% | 321 | 901 |
Personification is the worst on recall AND the most expensive. 53% recall at 548 prompt tokens vs 73% at 362 (GENERAL) or 87% at 321 (ULTRA-LEAN). Leaner + more depersonalized → recall up, tokens down. The exact opposite of the orthodoxy.
The hardest flaw tells the story
CONCEPT-1 — a len(text)//4 tokenizer heuristic that's a flawed
mental model — was the hardest to catch. On that one flaw alone:
- PERSONIFIED: 0/5 models caught it
- GENERAL: 2/5
- ULTRA-LEAN: 3/5
The personified prompts were uniformly blind to it. The depersonalized prompts found it more often than not.
Why ultra-lean isn't the answer
ULTRA-LEAN has the highest recall (87%) — but look at precision: it drops to 65%, meaning more false alarms. GENERAL is the sweet spot: high recall (73%) and the best precision (79%).
The recommendation isn't "as lean as possible." It's general + depersonalized + method-described. Describe the task neutrally, describe the method as steps, but don't tell the model who it is. That beats both personification (bias + cost) and ultra-leanness (precision loss).
What we think is happening
Personification injects training-data priors. "You are a code reviewer" pulls the model toward the average code reviewer in its training set — which is a defensible, conservative, slightly shallow pattern. It anchors on the persona instead of the task.
A depersonalized task description ("review for conceptual soundness by enumerating assumptions") gives the model the cognitive operation directly, without the persona baggage. The model doesn't need to be a reviewer — it needs to do the review. Those are different activations, and the latter performs better.
What this changed
Every production prompt in the harness is now a depersonalized task description. No "you are a planner" / "you are a coder" / "you are an adversary." Instead: "decompose this into tasks" / "implement this task" / "assume this is wrong, find where." The method is described as steps; the persona is absent.
Honest limitations
n=5 models × 3 flawed items. The absolutes are noisy at that size; the direction is consistent across all three metrics (recall up, tokens down, precision peaks at GENERAL). We haven't tested whether personification helps for generation tasks (writing code, drafting prose) — it may still help there. This probe is specifically about review/analysis tasks, where the cognitive operation matters more than the persona.