Meta Prompting, Reverse Prompting and Prompt Chaining

Study guide

Let the AI write and improve your prompts

Most people write a prompt, get a mediocre answer, and then rewrite the prompt by hand. This course teaches a faster route. You will ask the AI to improve your prompt before it runs the task, generate several prompt versions and keep the one that behaves best, and work backwards from a great output to recover the prompt behind it. Then you will chain prompts into staged workflows and use chain of density to produce short summaries that still carry the names, numbers and decisions.

The course is built for busy, non-technical professionals who use ChatGPT, Claude or Gemini at work. Every technique is explained in plain English, shown in a real job example, and paired with the mistake that usually breaks it. You do not need any technical setup. If you can paste a prompt into a chat box, you can use everything here. By the end you will have a small prompt library and a habit of treating prompts as reusable work rather than one-off typing.

What is a prompt framework?

A prompt is the request you type into an AI tool such as ChatGPT, Claude or Gemini. A prompt framework is a reusable structure for that request: a short checklist of the parts a good prompt contains, usually named with an acronym so it is easy to remember. Why it helps: the AI fills in whatever you leave out with generic guesses. A framework makes sure you include the parts that change the answer, such as who it is for, what you want to achieve and the format you need.

Frameworks make prompts repeatable: you fill the same parts each time, save the prompt as a template and share it with colleagues. They are not magic words: the value comes from the detail you put in each part. They help most when a request is complex or open to interpretation; a quick, simple question can stay short. There are well over sixty named frameworks, and most of them remix the same ingredients: role, task, context, audience, format, constraints and examples.

In this course you learn Meta prompting, which is a prompting technique from AI research: instead of slots to fill, it changes how the AI works through a task, and you can add it on top of any fill-in framework. It is best for: getting better prompts faster and turning one-off prompts into reusable workflows.

Let the AI improve your prompt first

Meta prompting means asking the model to write or improve the prompt before it does the task. You stay in charge of the goal, the model does the wording. The core instruction is simple: rewrite my request as a better prompt, ask me what is missing, then run it. That one message combines rewriting, gap checking and execution, so you get a stronger result without a second round of typing. Both OpenAI and Anthropic recommend using the model to optimise its own prompts, so this is a supported technique rather than a trick.

The gap check is the part people skip. When the AI asks what is missing, you supply the context it cannot guess, such as audience, tone or constraints. A good improved prompt usually has a role, a task, an audience, a format and constraints. If one is missing, say so before you run it. For example, a marketing manager asks for a launch email, the AI returns a rewritten prompt and asks who the audience is, and the manager answers that it goes to existing customers who already bought last year. That single detail changes the whole draft.

Meta prompting works in ChatGPT, Claude and Gemini because it only relies on instruction following, not on a special feature. The common mistake is accepting the rewritten prompt without reading it. The model may quietly change your goal, add a format you did not want, or drop a constraint you cared about. Read the rewrite, correct anything that drifted, then run it.

Automatic prompt engineering in everyday work

The automatic prompt engineer method, published by Zhou and colleagues in 2022, had a model generate and score many prompt candidates instead of relying on one hand written prompt. The everyday version needs no special tools. Ask for five versions of your prompt, run them on the same task, and keep the one that behaves best. Meta prompting here means writing instructions about how to write the prompt. You supply the role, the original prompt, the goal and the scoring rule, and the model supplies the candidates.

Score candidates on repeatability first. A prompt that produces the same sections and length every week beats a prompt that produces one brilliant answer and then drifts. OpenAI's metaprompting tip is to ask what phrases to add or remove so the behaviour you want appears more consistently. Consistency matters more than clever wording. A project coordinator who writes weekly status updates can test five prompt versions on last week's notes, pick the one that returns the same headings every time, and reuse it for the rest of the year.

Save the winning prompt. Automatic prompt engineering is a one off effort per recurring task, and the saved winner pays that effort back every week. The common mistake is judging candidates on one impressive output instead of on how they behave across several runs. Run each candidate at least twice on different inputs before you decide.

Reverse prompt engineering from a great output

Reverse prompt engineering starts from a finished output you admire and works backwards to the instructions that would recreate it. It is the mirror image of normal prompting, where you write instructions first and judge the result. The core request is simple: paste the output, say what it is, and ask for the prompt that would produce something similar. You are asking for runnable instructions, not for commentary on the text. If the reply is an essay about why the writing works, ask again for the prompt itself.

A recovered prompt is a template, not a finished tool. It still contains details from the original, such as brand names, topics and audiences, which you must replace with placeholders before reuse. The edit step is where the value is created. Strip the specifics, leave the structure, tone and format rules, then fill the blanks for each new job. One admired example can then serve many future tasks. Good sources include competitor newsletters, reports from other teams, job adverts and any writing that reads well. If you can see the pattern, you can recover the prompt behind it.

Reverse prompting pairs well with prompt libraries. Once a recovered prompt is cleaned and tested, it becomes a reusable entry you can pull out whenever a similar job appears. The common mistake is reusing the recovered prompt unchanged. It will keep producing the original company's tone, topic and examples, which is rarely what you want.

Prompt chaining into staged workflows

Prompt chaining splits a task into sequential prompts where each output feeds the next. You build the answer in stages instead of asking for everything at once. A typical chain runs research, outline, draft, review, final. Each stage is its own prompt, and the output of one becomes the input to the next. Anthropic notes that modern models handle most multistep reasoning internally. Chaining still helps when you want to inspect each step, because you can see which stage caused a problem.

Do not chain everything. If one prompt already gives a good answer, a chain just adds work. Chain when the task is long, checkable, or repeated often. Trim each output before pasting it into the next prompt. Raw output carries clutter, and clutter in the input weakens the next step. A consultant preparing a client report might run research, then outline, then draft, checking the outline before any drafting begins, which is far cheaper than fixing a full draft built on a wrong structure.

Keep the review step in every chain. It is the stage that catches wrong numbers, missing sections and tone problems before the final version goes out. The common mistake is chaining a task that was never a chain, then spending more time managing the stages than the task would have taken in one prompt.

Chain of density: summaries that keep the facts

Chain of density comes from research by Adams and colleagues in 2023. It rewrites a summary several times, adding key names and numbers while keeping the length the same. The result is a short summary that still carries the important facts. The everyday version of the technique is one sentence: make it shorter without losing the names, numbers and decisions. This tells the AI what to protect, which matters more than the word count.

Fix the length before you start. If you ask for a shorter summary, the AI will cut facts first. If you fix the length and ask for more density, the AI adds substance instead of removing it. Each pass should add one kind of detail. Pass one is plain text. Pass two adds names. Pass three adds numbers and decisions. The final pass is the densest version with no filler. Executive summaries and meeting notes are the best fits, because both are read by busy people who need names, numbers and decisions fast. A missing fact forces the reader back to the source document.

You can save the prompt as reusable slots: source, length, number of passes and facts to protect. Fill the slots once and you have a repeatable summary tool for any meeting or report. The common mistake is asking for a shorter summary without naming what must survive, then wondering why the budget figure disappeared.

Building your prompt library

A prompt library is only useful if each entry records when to use it, not just the prompt text. A prompt without context gets ignored or misused by anyone who finds it later. Store prompts as copyable text, not screenshots or images. If a colleague cannot paste it straight into ChatGPT, Claude or Gemini, it will not get reused. Sharing the best prompts turns individual wins into team assets. When everyone starts from the version that already worked, the team spends its time on the task, not on rebuilding prompts.

Models change, so a prompt that worked last year may behave differently now. Running a meta-prompt over the library asks the AI to review each prompt and suggest updates for current models, with reasons you can judge. A short review cycle works better than a big clean-up. Checking a few prompts every few months keeps the library alive without turning it into a project nobody has time for.

The whole course builds one habit, treating prompts as reusable work. Meta prompting, automatic prompt engineering, reverse prompting, chaining and chain of density all produce prompts worth saving. The common mistake is saving everything. Keep the prompts that solved a recurring problem, and let the rest go.

Frequently asked questions