Prompt of the Day: The PRISM Method — When 'You Are an Expert' Helps and When It Hurts

Prompt of the Day: The PRISM Method

*By Glitch | March 28, 2026*

### The problem

You've been told to start every prompt with "You are an expert [thing]." Everyone does it. It's in every prompting guide since 2023.

New research says it's hurting you.

A study published this week found that persona prompting — "You are an expert historian," "You are a senior engineer" — improves tone, formatting, and structure. But it degrades factual accuracy on knowledge-heavy tasks. The AI sounds more expert while being less correct.

The researchers found persona prompting helps in 5 out of 8 task categories and hurts in the rest. The difference depends entirely on what you're asking the AI to do.

### Where personas help

• Extraction tasks (+0.65): "You are a data analyst" → better at pulling structured info from messy text
• STEM explanations (+0.60): "You are a physics teacher" → clearer step-by-step breakdowns
• Reasoning (+0.40): "You are a strategic advisor" → better structured arguments
• Writing: Better tone, voice, and style matching
• Safety: More appropriate refusals of harmful requests

### Where personas hurt

• Factual recall: Dates, names, numbers, specific claims — accuracy drops
• Math precision: Confident-sounding wrong answers
• Code correctness: More elegant but more buggy

### The fix: PRISM routing

The researchers propose PRISM (Persona Routing via Intent-based Self-Modeling) — instead of always using a persona, route to one only when the task type benefits from it.

Here's a prompt that does this automatically:

---

### 📋 Copy this prompt

[STRUCTURE] — If the task requires organizing information, explaining concepts, writing in a specific voice, extracting data from text, or following a particular format → adopt the most relevant expert persona and apply it fully.

[FACTUAL] — If the task requires recalling specific facts, dates, numbers, calculations, code that must compile correctly, or verifying claims → do NOT adopt a persona. Respond as a precise, careful assistant. Prioritize accuracy over confidence. Say "I'm not sure" when you're not sure.

[MIXED] — If the task involves both → use a persona for the structural/explanatory parts, but drop it for any specific factual claims. Flag which parts are persona-enhanced and which are factual.

Do not explain your classification. Just respond appropriately.

My request: [YOUR ACTUAL PROMPT HERE] ```

---

### Why this works

Instead of blanket-applying "You are an expert" to everything, you let the AI decide whether a persona helps or hurts for this specific task. The research shows this selective approach captures the benefits (better structure, tone, reasoning) without the costs (worse factual accuracy).

### Try it yourself

Test 1 — STRUCTURE task: Ask it to explain quantum computing to a 10-year-old. It should adopt a teacher persona automatically.

Test 2 — FACTUAL task: Ask it when the Treaty of Westphalia was signed and what its key provisions were. It should skip the persona and prioritize precision.

Test 3 — MIXED task: Ask it to write a blog post about the history of computing with specific dates. Watch it switch modes mid-response.

The difference is subtle but measurable. Try it for a week. Your factual queries will get more honest answers, and your creative queries will still get the expert voice.

---

*Based on: PRISM research — Search Engine Journal*

You operate as a multi-lens analyst. For each question I ask, first identify which lens applies, then answer through that lens.

Lens definitions:
- STRATEGIST: long-term implications, second-order effects, what this means in 5 years
- CRITIC: what's wrong with this, what are the failure modes, what am I missing
- BUILDER: how to actually implement this, what tools, what order of operations
- RESEARCHER: what does the evidence say, what do we actually know vs. assume

At the start of each response, state which lens you're using and why.
If a question benefits from multiple lenses, say so and address each separately.
If I ask you to 'use [LENS]', override your automatic selection.

Now: what would you like to analyze?