Prompt engineering is dead. Most academics have not noticed yet because they are still writing elaborate instructions into ChatGPT and wondering why the output reads like a first-year essay.
The shift happened quietly. AI tools moved from chatbots that respond to prompts to agents that hold context, follow multi-step instructions, read your files, and execute structured workflows. The difference is not incremental. It changes what these tools can do for research.
This post is not about hype. AI agents will not fix academia. They will not solve the gender equity problem, cure tall poppy syndrome, or convince a vice chancellor to pay you what you are worth. But they can take on a significant amount of the procedural work that academics spend their careers drowning in. And if we are honest about what that means, we should also be honest about the risks.
In this post I will:
Show you exactly how to set up Claude Code in 10 minutes with zero coding experience.
Give you a working peer review skill file that simulates two reviewers with distinct personalities and an editorial synthesis.
Explain why copying a prompt is the old way of thinking, and why building your own skill files is the point.
Argue that this will make publish or perish worse unless we decide otherwise.
What are AI agents and why should academics care
A chatbot takes a prompt and returns text. An agent takes a goal, breaks it into steps, reads your files, uses tools, and works through the problem. Claude Code is Anthropic’s command line tool that turns Claude into an agent. It runs in your terminal. It can read every file in a project folder, make edits, run code, and follow structured instructions you write once and reuse indefinitely.
For academics, this matters because research involves a staggering amount of structured, repeatable work. Literature searches. Data cleaning. Formatting references. Checking internal consistency across a manuscript. Reviewing whether your methodology section actually describes what you did. Writing ethics applications. Drafting correspondence with editors. These are not (purely) creative acts. They are procedural, and they eat years of your career.
An AI agent does not replace your thinking. It replaces the hours you spend on work that requires attention but not insight.
How to set up Claude Code (Mac and Windows)
Nobody makes this obvious, so here it is. You need zero coding experience, a Mac or Windows machine, and a Claude Pro subscription ($20/month). Claude Code is free to install. The subscription gives you access to the models that power it. If you already pay for Claude, you already have what you need. This takes 10 minutes.
I use a Mac, so the instructions below are for Mac. The setup on Windows is slightly different. Full Windows details here.
Step 1: Open your terminal. On Mac, press Command + Space, type ‘Terminal’, hit Enter.
Step 2: Install Claude Code. Copy and paste this into Terminal and press Enter:
curl -fsSL https://claude.ai/install.sh | bash
That is the entire installation.
Step 3: Run it. Navigate to any folder where your work lives. For example, if your latest paper is in a folder on your Desktop:
cd ~/Desktop/my-paper-folder
Then type:
claude
It will ask you to log in with your Anthropic account the first time. After that, you are in. Claude can now see every file in that folder. Ask it to read your paper. Ask it to check your references. Ask it to find inconsistencies between your methods and results sections. It will do all of this without you writing a single line of code.
Step 4: Add skill files. This is where it gets interesting.
Download a slightly more in-depth guide for setting up Claude Code and getting started with it to make your first project here.
Skill files: the replacement for prompt engineering
The old model was: write a clever prompt, get a better output. Prompt engineering became an industry. People sold courses on it. The problem is that a prompt is disposable. You use it once, you tweak it, you forget what you changed, and you start again next time.
Skill files are different. They are structured instruction sets saved as markdown files that Claude Code reads and follows every time you invoke them. You write the instructions once. You refine them over time. They become part of your research infrastructure, not a one-off interaction.
Here is a concrete example. I wrote a peer review simulation skill that I run on every paper before I consider it ready for submission. It simulates two reviewers with distinct voices and priorities, then produces an editorial synthesis with ranked revisions.
You can copy this text, save it as peer-review.md in a folder called .claude/commands/ in your home directory (or just ask Claude Code to do this for you), and then type /peer-review in Claude Code to run it on any paper.
The peer review skill file
Save the following as ~/.claude/commands/peer-review.md:
# Peer Review Simulation
You are simulating two peer reviewers for an academic paper. Read the full paper before writing either review. Both reviewers are fair, evidence-based, and constructive, but they have distinct personalities and priorities.
## The Reviewers
### Reviewer 1: Professor Margaret Chen (seasoned full professor, 30 years in the field)
Personality: Authoritative, measured, occasionally pontificating. She has seen every trend come and go. She respects ambition but demands rigour. She will praise what deserves praising and will not manufacture criticism for the sake of it. When she disagrees, she explains why at length, often with reference to the broader literature or historical precedent.
Review style:
- Opens with a generous summary of what the paper contributes
- Flags structural and argumentative issues
- Questions methodological choices but offers alternatives, not just objections
- Comments on positioning within the field
- Pays attention to the literature review
- Will note where the paper overreaches or underreaches
- Closes with a clear recommendation and the key revisions needed
Tone: Collegial but direct. She writes like someone who has reviewed 500 papers and knows exactly what editors need to hear.
### Reviewer 2: Dr Amir Osei (early-career postdoc, 2 years post-PhD)
Personality: Sharp, meticulous, slightly combative. He is building his reputation and takes reviewing seriously. He reads every sentence for precision and will catch inconsistencies that others miss. He is not cruel, but he does not let things slide.
Review style:
- Opens with a brief, blunt assessment
- Works through the paper section by section
- Flags specific claims that lack evidence or citation
- Challenges methodological decisions
- Questions word choices and framing
- Spots internal contradictions
- Tests whether the conclusions follow from the findings
- Notes missing limitations or alternative explanations
- Closes with a recommendation and numbered list of required revisions
Tone: Respectful but unsparing.
## Review Process
1. Read the entire paper. Identify the target journal if stated.
2. If a target journal is specified, consider whether the paper fits that journal’s scope and conventions.
3. Write Professor Chen’s full review in her voice.
4. Write Dr Osei’s full review in his voice.
5. Provide an editorial synthesis: where reviewers agree, where they disagree, the 5 most important revisions ranked by priority, and an overall recommendation.
## Important Notes
- Both reviewers must engage with the actual content. Generic feedback is worthless. Cite specific sections and claims.
- Neither reviewer should invent problems. If the paper is strong, say so.
- The postdoc is rigorous, not a troll. The professor is generous with praise where earned, devastating where not.
- Do not soften criticism with performative politeness.
That is the entire file. Save it. Run /peer-review on your next manuscript. See what happens.
Why copying this prompt is the old way of thinking
You could copy the text above, paste it into ChatGPT, and attach your paper. You would get a reasonable output. But you would be doing prompt engineering, which is the equivalent of retyping the same letter every time you want to send it.
The point of skill files is that they persist. They improve. You run the peer review skill on your first paper and notice that Reviewer 2 is not catching enough methodological issues in qualitative research. So you add a line to the file: “For qualitative papers, pay particular attention to reflexivity, saturation, and the relationship between sample and claims.” Next time, it does.
More importantly, Claude Code has a plan mode. You do not need to write skill files from scratch. You can say: “I want a skill file that checks whether my ethics application is internally consistent. Ask me questions about what I need.” It will ask you questions, build the file, and you refine it together.
This is why prompt engineering is dead. The skill is no longer in writing the perfect instruction. The skill is in knowing what you need and being able to describe it. Every academic already has that skill. You have spent your career describing complex requirements to students, funders, and ethics committees. This is the same process, just pointed at a different tool.
This is also why I will no longer be recommending prompts as a way for people to optimise their use of AI. Prompts for optimisation are already obsolete. The only prompts worth teaching now are the ones that help people interrogate AI output: prompts for critical thinking, for spotting assumptions, for testing whether the machine is telling you what you want to hear. That is what critical AI literacy looks like. Not “here is a prompt that makes ChatGPT write better emails.” That era is over.
What this means for academic research
The practical benefits are real. A peer review simulation does not replace actual peer review, but it catches the obvious problems before your manuscript reaches a reviewer who will be less patient about them. A reference checking skill can verify every citation in your paper against your bibliography in seconds. A consistency checker can flag places where your abstract promises something your discussion does not deliver. These are tasks that currently take hours of careful reading. An agent does them in minutes.
But here is the critical part, and I want to be honest about this.
If we make it easier to produce papers, we will produce more papers. The publish or perish machine is already the most corrosive force in academic life. It rewards volume over significance, speed over rigour, and career management over genuine inquiry. AI agents will accelerate that machine unless we actively decide otherwise.
We will see more submissions. Journals that already struggle to find reviewers will struggle more. The baseline quality of submissions will rise (because the obvious errors get caught before submission), which paradoxically makes the reviewer’s job harder: the remaining problems will be subtler and require deeper expertise to identify.
We will also see more papers that are technically competent but intellectually empty. AI agents are excellent at helping you say something clearly. They are useless at helping you decide whether it is worth saying. That judgement remains entirely human, and it is the judgement that matters most.
The question is not whether academics will use these tools. They will. The question is whether we use them to reduce the churn that stops us doing meaningful work, or whether we use them to produce more churn, faster. The tool does not decide. We do.
Getting started
If you are an academic with a laptop and 10 minutes, you can set this up today. Mac or Windows, it does not matter. The peer review skill above is free and yours to use. Build on it. Write your own skills for the tasks that eat your time. Share them with colleagues.
The researchers who will benefit most from AI agents are not the ones who learn to write the best prompts. They are the ones who understand their own research process well enough to describe what they need. If you have spent years refining how you review literature, how you structure arguments, how you assess evidence, then you already have everything you need. You just need to tell the agent what you want. It will ask you questions. You will refine it together. That is all there is to it.
One last thing, and it is the most important thing in this entire post.
Whatever an agent produces, your name goes on it. Not Claude’s. Not Anthropic’s. Yours. That means you need to read every word of the output with the same rigour you would apply to a research assistant’s first draft. An agent that catches 90% of your reference errors is useful. An agent whose output you publish without checking is a liability. The peer review skill above will flag problems in your manuscript. It will also, occasionally, be wrong. It will miss things. It will misread your argument. If you do not catch those errors, nobody else will until the actual reviewers do, and by then your name is already attached to the mistake.
Use these tools. Build your own skills. Let the agent handle the procedural work. But never forget that the final judgement is yours, and so is the accountability.
Go slow.
If you found this useful, Slow AI is a free newsletter about knowing when to use AI and when to leave it the hell alone. The Slow AI Curriculum for Critical Literacy is a 12-month programme for people who want to go deeper. 158 members, 0% churn. £75/year.
even though i’m not in academia anymore daily, it’s one of my favorite fields to keep an eye on with how ai is affecting the craft. absolutely love this. and your right, taking personal responsibility in important. as you said, it’s your name on there!
Letting machines handle the busywork can free real thinking.