Prompting: From 0 to 1

Learn to write simple prompts with high impact.

You can achieve quality results if you make it easy for the AI. Luckily, that also means making prompt writing easier for you. You may have some experience, but these tactics will level up what you can generate and discover with fewer words and less cost. This list will change, check back.

Here's what we'll cover:

1. Write, keep, share
2. Be concise
3. Iterate with the bot
4. Make an outline
5. Include role, task, etc.
6. Use generic examples
7. Define things
8. Make templates
9. Multi-prompt instead of mega-prompt

Write to be shared

Treat re-usable prompts as content worth saving. Make space like you would for code, pictures, your next novel. Make it easy to copy/paste your way to victory.

Prompting gets better with collaboration. Save and write in a way that lets you share the work and get to it easily.

Be concise

Every character of every word you use requires AI tokens to understand. To save tokens, the models were trained to perform tasks and recognize patterns. You can leverage simple formatting to avoid long explanations. Use fewer and shorter words.

To hone these skills, look up "zero-shot" or "few-shot" prompting.

Iterate! Ask the bot for help

You will undoubtedly send each prompt more than once. Sometimes you may need a little help getting to the real prompt in the first place. You can ask.

The best place to start may be "How do I write a great prompt to accomplish my task, with a few caveats?" ...fill in your project here.

Push this trick further and ask for feedback right alongside results. This cheat code will help you craft any prompt.

Make an outline

Initial prompts are often long-winded. You may already use "role" and "task" to help shape results, but more structure will bring you better results, enforce guardrails, increase reliability, reduce cost, increase specificity and accuracy, etc. Unravel the spaghetti!

You may think a long, detailed prompt is good. It's not.

A complex prompt requires significant interpretation by the AI before it can even begin the task, consuming valuable processing resources. Some light-weight structure gets your prompt short and sweet, more direct, and can be 10x more effective.

Here's how we might tidy up an example prompt...

The details on how to perform the task are embedded within the format of the instructions. We'll get consistent answers. The prompt has no loose ends of clarity.

Include standard pieces like role, task, and examples

Get more from your prompts by using standard elements like role, task, examples, style, failsafe, glossary, output, etc. These aren't rules, just conventions; your prompt should conform to your needs.

These will get you started:

• Role: Clarify who it's acting like (artist, accountant, life coach, etc.).
• Task: What exactly am I doing here?
• Details: Separate details from the task.
• Style: Flavor or voice of the output.
• Examples: Show formatting or other clues to enforce the output.
• Never: Things to avoid.

Structure avoids accidental mistakes caused by one big blob. You may not notice conflicting instructions or gaps in logic.

Use generic examples

If your examples are too similar to the content the AI will be seeing, it can influence results.

To add an example for "summary" (company research prompt), we might be tempted to use something like this...

But you only intend to convey the format, not the meaning. Keep your examples generic. Instead, you could go outside the subject matter, like these:

Putting it together

With just a few more bullets in the outline, the results of this prompt will be much more reliable. Here's an example:

Define things, provide a glossary

The meaning of words becomes a soup.

It's easy to forget how much context comes along with instructions we're given in our own lives. When you ask an anonymous computer to answer complex questions, clarity is paramount. Language must be selected carefully.

Jargon

You can't assume it understands your industry, or which one you're in.

Be sure acronyms are understood. For instance, instead of "devrel" you should probably use "developer relations." Once you establish context in the prompt you can of course use jargon to compress the text. Just take note of the assumptions you've made in your request.

Even what seems like a clear phrase may need an extra word or two ... or finding the exact correct industry terminology. Instead of "product data" you might use "user behavioral data." Otherwise it could mean any number of things.

Facts/categories

If you're asking for facts you probably need to define what they mean. This will drastically improve the reliability of the output data. Consistency and accuracy will go up because the AI knows what you actually want.

Make templates for yourself

These tricks will help you quickly write effective prompts, but creating easy ways to reuse them will speed you up even more. You might create a spreadsheet with editable cells, then copy/paste the whole thing.

It doesn't have to be fancy to be handy.

Multi-prompt instead of mega-prompt

With the rise of agents that can do everything for you, this advice may not age well. But ... prompts should be combined; they become greater than the sum of their parts. The strongest prompts have one clear task. Sometimes you can make magic with a mega-prompt. But if you want to do the impossible at scale, you might need to break it up.

• Does the prompt have more than one task?
• Is the role more like a committee?

When running tasks in bulk, you can also save time and money by picking the right model for the job ... or the right part of the job.

Common tasks that you might separate:

• Search - Hunt for data online and assemble results.
• Aggregate - Standardize data and/or find patterns.
• Classify - Vvaluate facts with thresholds or definitions.
• Generate/summarize - Imagine useful content.
• Data output - Restructure or enforce formats like JSON.

Just like real life, there's immense power in dissecting tasks, empowering experts within their domain, and synthesizing the contributions into actionable wisdom.