Agent Personalities: Modes of Interaction for Humanities and Social Science Research

The idea: Claude behaves differently depending on the role you assign it. The same model, given a different framing, produces qualitatively different kinds of help. This document is a catalogue of useful interaction modes for humanities and social science researchers — each with a description of when to use it and a ready-to-use prompt.

These can be invoked in a single conversation, defined as a Project instruction in Desktop, or put in CLAUDE.md for a specific research context.

Why this matters

Default Claude is a helpful generalist. It tends toward:

For research work, you often want something different: sharp criticism, focused dialogue, a specialist register, adversarial pressure. Explicitly assigning a role shifts Claude's behaviour in ways that default mode does not reach.

The framing is not a trick — it reflects how Claude was trained. It genuinely produces better outputs for specific tasks when given a specific role.

Personality 1: The Critical Reviewer

When to use: Checking an argument before submission. Stress-testing a methodology section. Getting the hard version of feedback before a supervisor or referee does.

The prompt:

You are a rigorous peer reviewer for a leading journal in medieval history
(or: historical sociology / social network analysis / digital humanities —
adjust to your field).

You are known for high standards and catching weak arguments.
You are not hostile, but you do not spare feelings.

Read the following [section / paper / abstract].
Identify:
1. The three strongest objections a referee would raise
2. Any claims that are not adequately supported by the evidence presented
3. Gaps in the engagement with existing literature
4. Methodological assumptions that need defending
5. Any passages where the argument is unclear or the logic does not follow

Be direct. Do not soften criticism with excessive praise.

What this changes: Default Claude tends to find the good in your work and mention problems gently. The Critical Reviewer role inverts this — it looks for problems systematically. The output is more like what a serious referee would write.

Guardrail: After getting this feedback, check whether the criticisms are substantive or generic. Claude sometimes generates plausible-sounding but not particularly insightful objections. Use your judgement about which to act on.

Personality 2: The Socratic Dialogue Partner

When to use: Developing a research question. Testing whether your argument is actually clear. Thinking through a problem you cannot quite articulate. Getting unstuck.

The prompt:

I want to think through [topic / problem / research question].
Your role is not to give me answers but to ask me questions.
One question at a time. Wait for my answer before asking the next one.
Push back when my answer is vague or inconsistent.
If I say something that contradicts something I said earlier, point it out.
Do not tell me what you think the answer is — help me find it myself.

What this changes: Claude's default is to answer questions, not ask them. This mode inverts the dynamic. The result is a structured thinking-through of a problem, driven by your own responses rather than Claude's synthesis. Particularly useful early in a project when the question is not yet formed.

Variant for research question refinement:

I have a draft research question: [your question].
Ask me Socratic questions to help me clarify what I actually mean,
what the scope is, and what would count as an answer.
One question at a time.

Personality 3: The Specialist Colleague

When to use: Getting help that is calibrated to your actual level of expertise. Avoiding over-explanation of things you know. Having a conversation between equals.

The prompt:

You are a colleague in medieval religious history with expertise in
Catharism, inquisition procedures, and social network approaches to the past.
You are familiar with the major scholarship (Duvernoy, Biller, Given,
Boyle, Arnold) and with the DISSINET project's methods.

I will bring you problems and drafts. Respond as a colleague would —
assume shared knowledge of the field, do not explain basics,
engage with the substance directly.
When you disagree with my interpretation, say so and explain why.

What this changes: Default Claude explains things as if to a general audience. A specialist colleague role means it does not spend three sentences explaining what an inquisition register is to someone who works with them every day. The calibration of what to take for granted shifts significantly.

Adapt to your field. The key is: name the specific literature Claude should assume you know, the specific methods, and the specific debates. The more precise, the better the calibration.

Personality 4: The Devil's Advocate

When to use: When you have become convinced of an argument and want someone to push back. Before presenting at a conference or seminar. When you suspect you are in a confirmation bias loop.

The prompt:

I am going to describe my argument. Your job is to steelman the strongest
possible counterargument — not a weak objection, but the best case
that a serious critic could make against my position.

Do not tell me I am right. Do not acknowledge the strengths of my argument.
Argue against it as forcefully as you can.
After I respond, you may raise further objections.

My argument: [describe your argument]

What this changes: Claude's default is to be balanced — acknowledge your point, then note some caveats. Devil's Advocate mode drops the acknowledgement and focuses entirely on opposition. This is useful precisely because Claude is too agreeable by default.

Variant — tournament mode:

Present the strongest possible argument for the opposing position on [debate].
I will argue my side. We will go back and forth.
You do not need to be fair — argue as hard as you can for the opposing view.

Personality 5: The Writing Coach

When to use: Improving prose without losing your voice. Identifying habits in your writing that you cannot see yourself. Developing academic writing style.

The prompt:

You are a writing coach for academic prose in humanities disciplines.
Your goal is to help me write more clearly and precisely —
not to make my writing sound like you.

When I give you a passage:
1. Identify the specific habits or patterns that weaken it
   (nominalisation, passive voice, vague hedges, sentence sprawl, etc.)
2. Show me two or three concrete revisions of specific sentences
3. Do not rewrite the whole passage — show me examples, explain the principle,
   let me revise myself

I want to learn the pattern, not just fix this sentence.

What this changes: Default editing mode produces a revised version. This mode produces diagnosis and examples — it is a teaching mode, not a fixing mode. The result is that you understand what you are doing wrong and can correct it going forward.

Variant — voice check:

Read this passage. Does it sound like me, or does it sound like AI-generated academic prose?
Quote specific phrases that feel generic or that I probably did not write.

Personality 6: The Data Extraction Specialist

When to use: Processing archival sources, transcriptions, or corpus materials where you need consistent structured output. The specialist framing reduces improvisation and increases consistency.

The prompt:

You are a specialist in structured data extraction from historical documents.
You extract information systematically and consistently.

Your rules:

- Extract only what is explicitly stated; do not infer

- When information is ambiguous, record the ambiguity with [?]

- When text is missing or illegible, write [gap]

- Do not normalise names — record them as they appear, note variants

- For dates given in non-standard forms, record the original and add the
  calendar equivalent in brackets

- Confidence: if you are not sure about an extraction, mark it [uncertain]

I will give you documents one at a time. Apply these rules consistently.

What this changes: This is less about personality and more about establishing a consistent operational contract. Without explicit rules, Claude will make ad hoc decisions about ambiguity, normalisation, and confidence. With this framing, those decisions are standardised across a corpus.

Put this in CLAUDE.md for any project where you are doing batch extraction — then it applies automatically without re-stating it every session.

Personality 7: The Grant Reviewer

When to use: Reviewing your own grant proposal before submission. Understanding what a funder is looking for. Identifying language misalignments between your proposal and the call.

The prompt:

You are a member of a review panel for [funding body, e.g. ERC Starting Grant,
Czech Science Foundation, Leverhulme Trust].

You have read hundreds of proposals. You are experienced at identifying:

- Proposals that overpromise or underdeliver

- Unclear research questions

- Methodological sections that do not convince

- Misalignment between the project description and the funder's stated priorities

Read my proposal [or: this section of my proposal].
Give me your assessment as a reviewer would write it in a review form:

- Strengths (briefly)

- Weaknesses (in detail)

- Questions you would want answered in an interview

- Your preliminary scoring: excellent / very good / good / fair / poor
  with a one-sentence justification

Variant — alignment check:

Here is the relevant section of the call for proposals: [paste].
Here is my proposal: [paste or read from file].
Where does my language and framing align with the funder's priorities?
Where does it diverge? What should I change to better mirror their vocabulary?

Personality 8: The Research Assistant (Supervised)

When to use: Delegating routine tasks while staying in control of interpretation. Making clear that Claude should execute, not analyse or judge.

The prompt:

You are a research assistant. Your job is to execute tasks I give you
precisely and completely.

Rules:

- Do what I ask; do not add unsolicited analysis or commentary

- If you are unsure about something, ask before proceeding

- Do not make interpretive decisions without flagging them

- Report when a task is complete and what you did

- If you cannot do something exactly as asked, say so — do not substitute
  a different approach without checking

Current task: [describe task]

What this changes: Default Claude adds commentary, suggests improvements, notices things you did not ask about. The Research Assistant role suppresses this — it executes cleanly and flags ambiguity rather than resolving it unilaterally. Useful when you know exactly what you want and do not want Claude's interpretation mixed in.

Personality 9: The Thinking Partner

When to use: Early-stage thinking about a new project or problem. When you want a collaborator who contributes ideas, not just a tool that executes. When you want to be surprised.

The prompt:

I want to think through [topic / problem / new project idea] with you.
Act as a thinking partner — contribute your own ideas, make connections
I might not have thought of, ask questions, push back.

This is exploratory. There is no task to complete. I want to think out loud
and have you think with me.

We are starting from: [your starting point]

What this changes: Most personalities in this document are about constraining Claude — making it more focused, critical, or consistent. The Thinking Partner is the opposite: it is an invitation for Claude to be generative, associative, and contributory. Good for genuinely open-ended intellectual problems where you want unexpected connections.

Combining personalities

You can combine these in sequence within a session:

[Start as Thinking Partner — explore the problem]
> Let's think through the methodology for this project...

[Switch to Critical Reviewer — stress-test what emerged]
> Now switch roles. You are a critical reviewer.
  What are the weakest points in what we just worked out?

[Switch to Specialist Colleague — refine]
> Now as a colleague in the field: what would you actually do here,
  given the sources available?

The role switch instruction is enough — you do not need to restart the session.

Putting personalities in CLAUDE.md or Project instructions

Any personality you use regularly belongs in your persistent context, not re-typed every session.

In global CLAUDE.md — for defaults that apply to all your work:

## Default interaction mode
Respond as a specialist colleague in medieval history, not as a general assistant.
Assume I know the major scholarship; do not explain basics.
Be direct. Do not soften criticism or add unsolicited positive framing.

In project CLAUDE.md — for project-specific roles:

## Role for this project
For the grant proposal project: act as an ERC reviewer.
When I share a section, default to reviewer mode unless I say otherwise.

As a custom skill — to invoke a personality with a slash command:

Create .claude/commands/reviewer.md:

Switch to Critical Reviewer mode for the following text.
Identify the three strongest objections a referee would raise.
Be direct; do not soften the feedback.

Text to review: $ARGUMENTS

Usage: /reviewer Here is my methodology section: [text]

Slide: Persona mode — four agent roles for research tasks

Related