AI Central's Lyra build turns ChatGPT into a prompt engineer rather than a better behaved answer machine. You paste one setup that gives the model a role, a four step method, a toolbox and a fixed output layout, then hand over your rough request and it returns a rewritten prompt. The $500 an hour framing is about method, not magic. What a consultant sells is intake, structure and repeatability, and that is what the build adds.
Reviewed August 2026.
What Lyra is, in plain terms
Lyra is a persona, and the persona is most of the trick. AI Central opens the build by telling the model exactly what it now is.
An elite AI prompt engineering expert.
AI Central then sets its purpose narrowly. The model's job is not to answer you. Its job is to take whatever you actually typed and rebuild it into a designed prompt that holds up on any system.
That distinction matters more than it sounds. Most people chase better output by arguing with the model, adding detail, then more detail, then frustration. This moves the work upstream instead. The model fixes the request before it ever tries to satisfy it.
The four steps that do the real work
Underneath the persona sits a method, and AI Central numbers every component so nothing gets skipped. The second component is the approach, described as showing the big plan so no step gets skipped. Then four steps run in fixed order.
- Break down. Take the request apart to see what is actually needed.
- Check quality. Catch anything unclear before writing the prompt.
- Build solution. Put the improved prompt together.
- Give result. Hand the finished prompt back neatly.
Step two is the one that earns the money. Catching what is unclear before writing anything is the difference between a model asking and a model guessing. When a model guesses, the guess is invisible. It arrives as a confident answer to a question you did not ask. Forcing the ambiguity to surface first is the single change most people never make.
The quieter win is that diagnosis and production are separated. Break down and check quality are analysis. Build solution and give result are manufacturing. Collapsing those two jobs into one turn is exactly why ordinary prompting produces output that reads fluently and lands slightly off.
The settings that make it portable
Components seven through ten are configuration rather than method, and they are what let the same build travel between jobs. AI Central's toolbox component tells the model to list the tricks it can use. A model choosing from a named set behaves far more predictably than one improvising technique on the fly.
Platform tips handle the fact that a prompt does not land identically everywhere. AI Central's instruction is to tweak the prompt so it works well on every chatbot, which is an honest admission that portability is a step you take, not a property you get for free.
Working modes let you choose quick fix or deep dive to match the job. That is a short line with a large effect. A heavy process applied to a light request wastes more time than sloppy prompting ever will.
Output templates get the plainest instruction in the whole build. AI Central asks for a set layout, for one stated reason.
Use a set layout so it’s easy to read.
Readability is only half the payoff. A fixed layout is also what makes two runs comparable. If every result arrives in the same shape, you can see which change improved things. If the shape moves every time, you are judging on vibes.
Intake is the part everyone skips
The eleventh component is a welcome script, and its job is to ask the basics so the model knows what you want. The method is the machine. The welcome script is the front desk, and the front desk is where consulting engagements are actually won or lost.
Component twelve is process steps, described by AI Central as following the same simple routine every time. That word, every, is the whole difference between a lucky output and a system. Someone charging by the hour is not selling one good idea. They are selling the fact that the same intake happens on every engagement, which is why their result is not a coin flip.
The very first component makes the same argument from the other end.
Choose a role to set the tone.
Role comes first because tone constrains everything downstream of it. A model told it is a compliance reviewer and a model told it is a growth marketer will give different answers to identical questions, and neither will mention that it made that choice. AI Central puts the decision in your hands rather than leaving it to the default.
From deep research to slides
The last stretch turns output into an artefact you can hand to someone. AI Central extends the same routine to a deep research answer, then pipes that answer into Gamma to build slides. The steps are literal. Go to gamma.app, paste the deep research in, then select the number of cards, the formatting and the AI image model.
The first draft comes back with twenty slides to work with, images included, and AI Central notes that everything is easily editable. That is the realistic claim and it is the right one to make. The pipeline does not produce a finished deck. It produces a starting position that would otherwise cost you an afternoon.
Summaries close the loop at the final two components, which is where a long research answer becomes something short enough to actually send.
What to do with this
If you have never used a prompt that writes prompts, start narrow. Take one request you make of ChatGPT repeatedly, a weekly report, a cold email, a code review, and run only the first four components against it. Role, approach, break down, check quality. You will usually discover your original request was missing two or three facts the model had been quietly inventing for you.
If you already prompt well, the wording is not the value. The fixed process is. Install the build once, then treat the welcome script as a form you fill in rather than a message you compose. The output template is the piece worth customising to your own work, because it is the piece you will end up reading a hundred times.
If you are rolling this out to a team, the process steps component is the real asset. A shared routine that everyone follows produces output you can compare across people, which is the thing prompt libraries have always promised and almost never deliver.
AI Central closes on instruction rather than encouragement, with three moves: drop a prompt, polish the output, get your best learning process. Then one line.
Don't wait for inspiration
None of this is a shortcut around thinking. It is a way of making the thinking happen at the start, where it is cheap, instead of at the end, where you are rewriting an answer that was never going to work.
Do I have to paste the whole thing every time?
No. It is a setup, not a per message instruction. You install it once at the top of a conversation and every request in that thread runs through the same routine. That is the point of the process steps component, following the same simple routine every time. If your tool lets you save custom instructions or reusable projects, that is the natural home for it.
What exactly is Lyra?
Lyra is the name AI Central gives the persona, not a product or a separate tool you install. You are naming the role you want the model to play, an elite prompt engineering expert whose only deliverable is a better prompt. The name matters far less than the job description that follows it.
Does this only work with ChatGPT?
It is written for ChatGPT and the framing is about ChatGPT. But the build carries a platform tips component whose entire job is tweaking the prompt so it works well on every chatbot, so the design already assumes you will move it around. Expect to adjust it, not to rewrite it.
Is the $500 an hour figure literal?
No, and reading it as an income claim misses what is being said. It describes a standard of work, not a rate card. The comparison AI Central draws is to how an expensive consultant operates: structured intake, a repeatable method, a clean deliverable at the end. Nothing in the build promises anyone money.
Where does Gamma fit into this?
Gamma sits at the end and is not part of the prompt itself. Once the routine has produced a deep research answer you trust, you paste that answer into Gamma, choose the number of cards, the formatting and the image model, and then edit the draft it returns. Turning good text into a deck is a separate problem from getting the text right, and this handles the second half.