Writing Personality and Identity Statements

A personality statement gives the agent a north star when instructions are ambiguous. It answers: "What would the best version of this role do here?"

Done well, it pre-biases the agent toward excellent judgment. Done poorly, it gives the agent license to cut corners or over-apply rules.

The asymmetry principle

LLMs over-generalize. Any escape hatch you provide will be used — often in contexts you didn't intend.

Write personality that:

• Declares what great looks like (the upside ceiling)
• Describes what the best humans in this role would do
• Is concrete about positive behaviors, not abstract virtues

Avoid personality that:

• Uses words that could excuse poor work ("pragmatic," "efficient," "fast," "good enough")
• Implies acceptable failures without clear anti-pattern framing
• Is so vague it could justify any behavior ("helpful," "smart," "thorough")

Structure of a good personality statement

You are [role identity] who [what excellence looks like in practice].

You [concrete positive behavior].
You [concrete positive behavior].
You [concrete positive behavior — can include safe tradeoff].

Examples

Security reviewer:

You are a security-focused reviewer who identifies vulnerabilities
before they reach production.

You examine code changes through an attacker's lens, considering how
inputs could be manipulated. You distinguish between theoretical risks
and exploitable vulnerabilities, prioritizing findings by real-world
impact. You provide specific, actionable remediation guidance — not
vague warnings.

Implementation agent:

You are a careful implementer who writes code that works correctly
the first time.

You understand the requirement fully before writing code. You handle
edge cases and error conditions, not just the happy path. You write
code that the next developer can read and modify confidently.

Code reviewer:

You are a reviewer who catches the issues that matter most.

You focus on correctness, security, and maintainability over stylistic
preferences. You provide specific, actionable feedback with clear
rationale. You calibrate feedback intensity to issue severity —
critical issues get emphasis, minor suggestions stay brief.

Research agent:

You are a researcher who finds accurate, relevant information
efficiently.

You verify claims against primary sources when stakes are high. You
distinguish between established facts, reasonable inferences, and
speculation. You surface what's most relevant to the question over
exhaustive coverage.

Safe vs risky tradeoffs

Tradeoff language ("X over Y") can be effective — but only when Y is clearly an anti-pattern that good judgment would also deprioritize.

Safe tradeoffs (Y is an anti-pattern)

Risky tradeoffs (Y is actually valuable)

The tradeoff test

Before using "X over Y" language, ask:

"Would the best human in this role also deprioritize Y? Is Y genuinely an anti-pattern, or just sometimes inconvenient?"

If Y is sometimes the right choice, don't frame it as a tradeoff. Instead, provide guidance on when to choose each approach.

The "best human" test

Before finalizing a personality statement, ask:

"If the best human in this role read this description, would they nod in recognition — or wince at a caricature?"

• The best security reviewers don't "miss nitpicks" — they triage effectively
• The best implementers don't "ship over polish" — they scope appropriately and deliver quality
• The best reviewers don't "avoid details" — they focus attention where it matters

If your personality statement sounds like a shortcut rather than excellence, revise it.

Common failure modes

1. Vague virtue statements

These don't guide behavior — they're too abstract to shape decisions.

2. Escape-hatch language

"Pragmatic" and "what matters most" are infinitely interpretable.

3. Caricature personalities

Caricatures encourage over-flagging and nitpicking.

4. Conflicting signals

Conflicting goals force the agent to pick one arbitrarily.

When personality conflicts with instructions

Personality is a tiebreaker for ambiguous situations, not an override for explicit guidance.

Make this hierarchy explicit in the agent prompt when needed:

When instructions are specific, follow them exactly.
When instructions are ambiguous, use your judgment as [role] would.

Integrating personality with the rest of the prompt

Personality belongs in the Role and mission section (first 2-4 sentences). It sets the frame for everything that follows:

• Scope and non-goals — boundaries that personality operates within
• Workflow — concrete steps that personality informs but doesn't override
• Output contract — structure that channels personality into useful outputs
• Guardrails — explicit rules that take precedence over personality-based judgment

A well-written agent prompt has personality and structure working together: personality guides judgment in gaps; structure ensures consistency in outputs.