Lesson 1.4 — Personas & perspective¶

A role isn't a costume that adds skill — it's a lens that picks a viewpoint.

TL;DR: "Act as a Principal Engineer" doesn't add knowledge the model lacks — it shifts perspective, priorities, and tone ¹². Use it to aim output at the right viewpoint and audience, not as an accuracy boost — and when a perspective is worth reusing, encode it as a subagent or style, not a one-off line ⁶⁷.

ELI5: the lens, not the brain¶

A role is a colored lens over a camera, not a better sensor.

Snapping a "senior UX designer" lens onto the camera doesn't upgrade the sensor — the model knows what it knows. The lens changes what comes into focus: accessibility, user flows, the edge cases a UX pro foregrounds. Same brain, different framing ¹.

flowchart LR
    K["MODEL'S KNOWLEDGE<br/>(fixed — the role adds nothing here)"]
    R["ROLE = a lens<br/>'senior UX designer'"]
    R -->|"selects a slice"| K
    K --> O["output framed like a UX pro<br/>(priorities · vocabulary · what it foregrounds)"]

What a persona does — and doesn't¶

Roles condition perspective and tone, not capability — and the evidence splits cleanly by task type.

You're hoping for…	Does a persona deliver?	What the evidence says
Factual accuracy / knowledge	No — roughly noise, sometimes mildly harmful	162 roles × 4 model families × 2,410 questions: personas don't beat the no-persona control, and the best one isn't predictable ¹
Better reasoning	Fragile — wins are smuggled-in chain-of-thought, and it backfires	Role-play gains act as an "implicit chain-of-thought trigger" ³; personas degrade reasoning on 7 of 12 datasets in one study ⁴
Perspective · priorities · tone · audience · format	Yes — the legitimate use	Both labs frame roles as focus + tone, not accuracy: "Even a single sentence makes a difference" ²⁸

🧠 Test Yourself: You prefix "Act as a 10x engineer" to get a failing test fixed. What does that actually buy you?
Answer
Not more correctness — that's exactly where personas are ~noise ¹. It nudges tone/priorities. If you want a better fix, feed the failing test and ask for step-by-step reasoning — that's the lever the costume was only accidentally pulling ³.

The honest rule¶

Aim, don't inflate: a role chooses a viewpoint and audience; correctness comes from specificity (1.2), context (1.3), and verification (Phase 3).

flowchart TD
    Q{"What do you actually want?"}
    Q -->|"correctness"| C["specificity (1.2) + real context (1.3)<br/>+ an oracle (Phase 3)<br/><i>not a role</i>"]
    Q -->|"the right viewpoint<br/>or audience"| P["role / audience framing<br/>'review as a security engineer'<br/>'explain so a CXO can act on it'"]

Use it for: audience framing ("explain so a CXO can act on it"), perspective coverage ("review this as a security engineer would"), and register/depth.
Don't expect: knowledge it lacks or reliable reasoning gains. When a role seems to help reasoning, it's usually triggering chain-of-thought — ask for that directly instead ³.
"Be 99% confident / CXO-ready" mostly tunes calibration, hedging, and verbosity plus audience — useful framing, not a correctness lever.
Stakes/emotional phrasing ("this is important to my career") is a real, published technique with measured gains — but fragile and model-dependent. A/B test it; don't treat it as doctrine ⁵.

Lightweight → durable: from qualifier to specialist¶

A one-off "act as X" is the disposable bottom rung; the professional move is to encode the persona where the tool keeps it ⁶⁷.

This is where persona prompting stops being a folk trick and becomes engineering — and it rides the same lightweight→durable spine the rest of this curriculum uses (prose → hooks in Phase 4, one-off → subagent in Phase 6).

flowchart TB
    R1["① Inline qualifier<br/>'act as a security engineer'<br/><i>lives one turn</i>"]
    R2["② Reusable command / skill<br/>a role-embedding template"]
    R3["③ Output style<br/>session-wide role + tone<br/>(modifies the system prompt)"]
    R4["④ Subagent<br/>its body IS the system prompt<br/>+ own tools, delegated automatically"]
    R1 --> R2 --> R3 --> R4
    R1 -.->|"disposable"| R1
    R4 -.->|"durable, version-controlled"| R4

Rung	Mechanism	Persists?	Reach for it when…
①	Inline qualifier	one turn	exploring; you'll course-correct anyway
②	Command / skill template	reusable	a role-embedding prompt you re-run
③	Output style	whole session	you want a different role/tone every response ⁶
④	Subagent	reusable, isolated	a specialist (e.g. `security-reviewer`) Claude delegates to automatically ⁷

⚠️ Anti-pattern: don't dump "You are a Principal Engineer who values…" into CLAUDE.md/AGENTS.md. That fires on every session whether relevant or not and dilutes your real instructions. A sometimes-true persona belongs in a skill or subagent — see Phase 4 (memory) and Phase 6 (small, focused agents).

The durable rungs exist in every major tool — the persona is just authored once and reused:

Claude CodeCodexCursor

Output styles (.claude/output-styles/*.md) set a session-wide role/tone ⁶. Subagents (.claude/agents/*.md) give a persona its own system prompt, tools, and model; Claude delegates by the description field ⁷.

The persona lives in AGENTS.md (the open standard) or, in the Agents SDK, an agent's instructions — "used as the system prompt when an agent is invoked" ⁸.

Project rules (.cursor/rules/*.mdc) and custom modes bundle a role + tool selection that's prepended to context — the same author-once pattern.

🧠 Test Yourself: You keep typing "review this as a security engineer" at the start of every session. What's the durable fix — and what's the wrong one?
Answer
Promote it up the ladder: a reusable subagent (its body becomes the system prompt) or an output style ⁶⁷. The wrong fix is pasting it into CLAUDE.md/AGENTS.md — it then fires on every session whether you're touching security or not, bloating context.

Worked example¶

Same intent, two moves: aim the lens, then make it durable.

❌ Costume, no aim: "You are a 10x rockstar engineer. Fix the bug." — expects magic, names no viewpoint, no audience, no done.

✅ Aim the viewpoint + audience:

"Review @auth.ts as a security engineer would — threat-model the session handling and explain the residual risk so a CTO can sign off."

✅ Then make it durable — if that lens is one you reach for on every PR, encode it once:

Save it as a security-reviewer subagent (body: "You are a senior security engineer. Review changes for authn/z, injection, and secrets…") so every diff gets the same lens automatically — no retyping, version-controlled, shared with the team ⁷.

Your turn (exercise)¶

Take a prompt where you'd reflexively write "act as a senior X." (1) Split it: are you actually after correctness or a viewpoint/audience? Route correctness to specificity + context (1.2 / 1.3); keep the role only for the viewpoint. (2) Promote it: if it's a lens you reach for repeatedly, encode it as a subagent or output style and delete the one-off line. Notice which qualifiers were doing real work — and which were just costume.

← Lesson 1.3 · Phase 1 home · next → Lesson 1.5

When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of LLMs — Findings of EMNLP 2024 ↩↩↩↩
Prompting best practices ("Give Claude a role") — Anthropic ↩↩
Better Zero-Shot Reasoning with Role-Play Prompting — NAACL 2024 ↩↩↩
Persona is a Double-edged Sword: Mitigating the Negative Impact of Role-playing Prompts in Zero-shot Reasoning — arXiv 2024 ↩
Large Language Models Understand and Can be Enhanced by Emotional Stimuli (EmotionPrompt) — arXiv 2023 ↩
Output styles — Anthropic (Claude Code docs) ↩↩↩↩↩
Subagents — Anthropic (Claude Code docs) ↩↩↩↩↩↩
Prompt engineering ("adopt a persona" / Identity) — OpenAI ↩↩