I found one sentence that supercharged the quality of my AI output. Before I tell you what it is, take a look at these two websites generated by nearly identical prompts.
I’m guessing you liked the design of Marginalia more than Folio. Here are the prompts I used for each:
Build a responsive reading-list app. Users can add books, mark them as read, and filter between read and unread. Save the list across refreshes. Make it work on desktop and mobile.
You are the best UX/UI designer in the world.
Build a responsive reading-list app. Users can add books, mark them as read, and filter between read and unread. Save the list across refreshes. Make it work on desktop and mobile.
Just adding that one sentence telling the model it is the best in the world at something made a large difference. You can see a difference in the font selection, the layout of the page, the use of color, spacing between sections, and hierarchy in the UX.
Both used GPT-6 Astra Ultra, in empty folders, built at the same time with no knowledge of each other. I checked the traces to be sure there was no contamination. Small design choices changed across reruns, but the extra sentence won every time.
I’ve been doing this for most of my prompts for a while now. I’ll tell Claude or Chat that it’s the best designer in the world, an expert software architect, the world’s most renowned writer, the most intelligent investor, etc. I’ve found big improvements in the quality of code, number of bugs that are caught, and the clarity of the writing produced.
So I tried to automate it
I’ve always found adding this tedious, so I decided to make a skill to save some keystrokes and do less babysitting. A skill is just a set of instructions your agent can reuse. You can find it on GitHub, and I’ve included installation instructions at the end. As it turns out, creating the skill was a lot more involved than I anticipated.
For my first attempt, I asked Astra to generate a skill that could reliably produce the difference I saw between Folio and Marginalia. I told it to frame the model as the best in the world at whatever expertise the task at hand required. I gave it both prompts and let it cook. It failed.
Astra decided to make an epically complicated set of instructions, and the prompt I wanted got buried in all the instructions. I took a look at the reasoning trace and the instructions were completely ignored as a result of all this complexity.
But I was too lazy to write the skill myself. So I had Astra take another crack at it, this time telling it not to stop iterating until it was able to show a demonstrable improvement from using the skill. I was actually quite impressed at the loop it created; there were multiple subagents to edit the skill, test the skill, and review the outputs from each prompt. It created separate containers to run each test. I was skeptical about hill climbing for a simple prompting task, but let it go. After 20 minutes I got a notification the top of the hill had been reached, and the test result was perfect.
However, there was just one problem: buried in the skill was a prompt designed for this specific example. Here’s what Chat had to say when confronted: “You’re right—I overfit a general-purpose skill to our UI test case.” So we removed the task specific language, and the skill became useless again.
What was the model paying attention to?
I had two guesses for why my “best in the world” framing worked:
- The higher standard pushed the model to run more iterations
- The UI/UX designer role emphasized the importance of getting design right
I looked back into the reasoning traces from my runs that produced better results and found evidence of both. There was both an added design phase and more reasoning spent overall. The skill I built was choosing to call the role “front end engineer” and thus extra time was spent on robustness of the application and edge cases that could cause bugs. Both roles were important, but the AI was choosing to focus on technical correctness and meeting the exact requirements of the original prompts. How could I push it farther?
So I decided to ask two questions:
If every literal requirement were satisfied, how could the result still fail its intended use?
What would make the audience prefer one equally correct result over another?
What made it work
The result was the following four step process, defining specific roles for the AI to play and then pushing it to achieve a higher standard of output:
- Define distinguishing strengths before production. What would make the result stand out? LLMs need to be focused on something to produce the best output. Just saying “build the best website” won’t perform as well as saying “design the best UX for this website, pay attention to spacing, font, and colors in the UI, and test it as a user would to ensure there is as little friction as possible”. Manually writing out the last prompt is tedious, but a prompt to force the LLM to make a meta assessment of the roles required before proceeding seems to produce similar results. A human employee would operate similarly; training on what to look for and how to structure thoughts makes somebody more productive. We need to define the lens through which the AI will view the task, but to make the skill generalizable, we need the AI to decide what that lens should be. We will inevitably leak some performance by ceding this responsibility to the AI (which may not have the user’s exact intentions in mind) but my results testing the skill so far have suggested that we get pretty close.
- Calibrate the standard with a reference. Inspect a strong, relevant reference or create a small concrete alternative. This gives “excellent” something tangible to compare against. After achieving a perfect test result, the AI finds a relevant reference or creates an alternative. For a website, that could be trying a new layout. With an alternative to compare against, “make it excellent” means a lot more.
- Review craft separately from correctness. Ask: “Where is this merely adequate, and which specific refinement would most improve the audience’s experience?” Assess the whole composition and how its parts work together. This makes outputs more coherent, whether it’s the structure of an essay or the overall theme of a website.
- Refine and retain the stronger result. More edits don't automatically mean better work. The skill preserves the previous version, compares substantive changes, and keeps the stronger result. If an edit makes things worse, it gets rolled back. Otherwise, all that extra effort can leave a mess behind.
The final result uses colors and fonts nicely with visually appealing spacing. It reduced some of the busyness I didn’t like in Marginalia (something I verified was an intentional edit in the reasoning trace) but did a better job at using color and element design in the sidebar and the selectors than Folio.
Why isn't this built into the tools already?
Every harness like Codex and Claude Code has to balance quality, speed, and cost. At some point, the agent has to decide the work is good enough.
But “good enough” still leaves a lot of room for improvement. Regardless of the harness I use, they all stop earlier than I’d prefer. I’d rather use the extra tokens to get a better result. And importantly, “better” needs something concrete behind it. What does a strong result look like? Which part of the result is still merely adequate? And did the latest change actually improve anything?
Try it yourself
I’ve packaged this whole process into Prompt Lab.
To install it, paste this into Codex:
1$skill-installer Install the skill from https://github.com/coltonconley/prompt-lab/tree/main/skills/prompt-lab
If the skill doesn't appear after installation, restart Codex. Then add it before your normal task:
1$prompt-lab2[Your task, constraints, audience, and desired result]
You don’t have to pick the expert roles or write out the review process. Include the context and constraints you normally would, especially anything important about who the result is for. Then let it cook.
My suggestion: try it on something you’ve already asked AI to do and weren’t happy with. Run the same task in fresh chats, with and without the skill, using the same model and settings. Compare the actual outputs and whether the improvement was worth the extra time.
I’m particularly interested in examples beyond website design. Writing, coding, analysis, whatever you’re actually using AI for. If it works for you (or if it doesn’t), reply with your prompt and the before and after. The more examples I have, the better the skill can be.
