I have been writing prompts for a very long time. I am sure a lot of us.
Then one weekend I asked a different question: what if AI writes the prompt, and I only decide what good looks like?
I tried it on something this weekend, a daily briefing agent that reads our Jira and Confluence activity and emails a summary for a business and technology audience. It took about six hours in total. It worked. Of-course a simple agent. My main aim was to make sure AI does its own prompting and I am not prompting it , at-least 95% less than what I used to.
So can AI prompt by itself (I only decide what outcome I need)? It was quite intriguing.
But the interesting part is where it went wrong first.
Today in the newsletter:
What the agent got wrong
The line that mattered most
The Prompt Tax
What the research says
The first version was fine, and useless to me
The agent did exactly what I asked, and that was the problem. It does not care about outcomes (sometimes generates the outcome you want, but with no brains behind it). It follows the prompt, finishes the job and has no idea whether the newsletter is any good.
My first output was tuned for technical people. I am in product and business, so for me it was mostly noise. I had never told the agent who I am, or what I do with the briefing once I read it (i.e outcome).
So most of the real work was spelling out what a high-quality briefing looks like for me and what I will do with it, and how it should help someone in product and business make decisions.
How I built it
I started with one line: “Review important Jira activity from the last 24 hours and email me the relevant changes.” That went into a prompt file that Claude Code picks up, and I ran it through Windows PowerShell, read what came out, and asked another LLM to rewrite the file. Four or five rounds of that. The LLM wrote most of the wording. I decided what mattered.
Along the way the prompt learnt to group tickets into business topics instead of one incident per ticket, to put customer, payment, revenue, security and compliance risks first, to explain technical terms in plain English, and to say what it could not verify.
Then I noticed that a lot of important change lives in Confluence, not Jira, so I added architecture pages, security and governance documents, runbooks, release plans and so on (and told it to ignore formatting-only edits). Today it picks five topics from either source and adds a management watch list, some positive progress and a note on data-quality limits.
I call this the Judgment Loop:
you define good, AI drafts the prompt, you run it, you review the output, AI fixes what failed, and you only automate once you trust it.
I am still between “run it” and “trust it”, which I will come back to.
“Do not invent facts”
This is the most important line in the whole prompt, and it exists because of a specific habit. The agent reads a single Jira ticket and draws a conclusion. “Highly critical”, it says. Maybe it is. Maybe it is not.
It does not know the wider business context, and it will happily sound confident anyway.
So now, if it cannot confirm customer impact, ownership, approval status or deployment status, it must say so.
The Prompt Tax
Let us give the pain a name. The Prompt Tax is the time you spend rewording the same instruction to make up for what the AI does not know about your task. PRD review, weekly update, competitor scan: every time you rewrite your framework, you pay it again.
The way out is to give the AI your context ONCE (the purpose, the audience, the decision the output should support, your framework, the format, examples of good and bad), and ask it to turn that into a working prompt. You review it and reuse it.
This is also why prompt libraries never stuck for me.
A library is somebody else’s context, and it cannot know your audience or your definition of quality.
What the research says
I wanted more than my own opinion, so I went looking.
In the 2022 APE paper, LLM-written instructions matched or beat human-written ones on 19 of 24 tasks.
VMware researchers found AI-optimised prompts did as well or better than hand-written “positive thinking” prompts in nearly all tested cases.
A 2025 study found LLM-written prompts lined up more closely with human labels, and varied less, than prompts from 15 human experts.
And there is research suggesting persona lines like “act as an expert” can sometimes reduce factual accuracy, which is awkward for anyone who still opens every prompt that way.
These are research tasks, not your PRD, so take them as direction, not proof.
What I would recommend to you
Pick one task you repeat every week. Tell the AI who the output is for and what good looks like, ask it to write the prompt, and ask it to question YOU about anything it is missing. Then read the first output with suspicion, the way you would read a new joiner’s first report.
Sources
Zhou et al., Large Language Models Are Human-Level Prompt Engineers (2022): https://arxiv.org/abs/2211.01910
Arabzadeh and Clarke, A Human-AI Comparative Analysis of Prompt Sensitivity in LLM-Based Relevance Judgment (2025): https://arxiv.org/abs/2504.12408
PromptHub, One Size Does Not Fit All (on VMware research): https://www.prompthub.us/blog/one-size-does-not-fit-all-an-analaysis-of-model-specific-prompting-strategies
Search Engine Journal, “You Are an Expert” Prompts Can Damage Factual Accuracy: https://www.searchenginejournal.com/research-you-are-an-expert-prompts-can-damage-factual-accuracy/570397/
PRISM paper on persona prompting and accuracy: https://arxiv.org/abs/2603.18507
Anthropic, Prompt Engineering Best Practices: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices