|
|
If I told you to be specific when you’re prompting AI, how would you interpret that? |
Because there are two different things you can be specific about when you talk to AI, and you’ll have an upper hand if you’re aware of this distinction. |
You can be specific about the path: the steps you want the AI to take, in the order you want them taken. Or you can be specific about the outcome: the result you're actually trying to get to. |
Many folks -- especially those with some sort of business background that had them writing standard operating procedures (SOPs) at some point in their career -- default to spelling out the steps. |
This feels rigorous (you're being specific, after all). But handing the model your step-by-step plan caps the result at the plan you already drew: it can only execute inside the boundaries of the constraints you provided (in theory — you can’t completely keep AI from coloring outside the lines). |
The shift I want to walk you through today is simple: be specific about the outcome you want to achieve, let the AI outline all the available paths to get there, and then use your better judgment to choose the path. |
Notice that you're not being less specific. The specificity just moves off the steps and onto the outcome. Also notice that last bit: you choose. Your judgment is the most important piece of this entire puzzle. |
So today I want to break down: |
How outcome-first prompting works How to start prompting this way today (with before and after examples) Why it’s possible today (plus, a reminder that context is queen)
|
—@dharmesh |
|
|
How Outcome-First Prompting Works |
In matters where you aren’t supremely confident in your own domain expertise, it makes sense to let someone else lay out all the possible options. As we all learned in school, multiple choice is easier than trying to produce the right answer from scratch. |
Prompting like this leans into the expertise of the model and exposes new paths to your desired outcome that you may have never considered. |
Here’s how it works: |
1. Be specific about the outcome. You're not describing how to do the work anymore, but you are describing what done looks like. Bonus points for what done well looks like. Include the deadline and any other real constraints you’re working within. Every bit of detail you used to spend on a step-by-step plan goes here. |
2. Ask for the different paths. Your wording here matters. If you ask “what are all the ways I could get there?” you’ll get quite a few paths. “What are the 3 most common paths to achieve this?” narrows things down. “What would a top performer within Industry X do to achieve this?” adds perspective expertise. |
3. You make the call. The model's job is to widen your option set. It doesn't get the final call, because it doesn't know what you know: your team, your commitments, your appetite for risk. |
This style of prompting is all about surfacing different paths that all lead to the same place: the specific outcome you defined. |
Now let me show you some examples of how you can implement this. |
|
|
How to Start Prompting This Way |
Each example below starts with the prompt you'd naturally write. Nothing wrong with it -- it'll get you a perfectly good answer. But the approach is already decided, so the most AI can do is polish your first idea. The outcome-first version lets your first idea compete with paths you hadn't considered. |
The lead magnets. Say you ask for "5 lead magnet ideas". The outcome-first version: "I want to double signups in 90 days. Here's my current traffic breakdown and top sources of subscribers. What are all my options?" |
The slide deck. Say you ask for “an outline for a slide deck.” The outcome-first version: “I need my boss to approve a $3,000 tool budget, and she’s skeptical of new software. What are the top 5 strongest ways to visually communicate the value of adding these tools?” |
The email flow revamp. Say you ask for “a new email for my welcome series”. The outcome-first version: “I want a new subscriber to go from ‘who is this?’ to ‘I’d happily pay this person’. Map out a few different welcome sequences that could get them there, highlighting the tradeoffs of each.” |
There is value to be found in asking open-ended questions even if they aren’t necessarily outcome-first: |
The customer review analysis. Instead of asking “Sort these 40 customer reviews into positive or negative categories”, ask “What are 3-5 of the most practical insights we can pull from these reviews? Include a short paragraph on how we can maximally apply that insight.” |
By keeping our prompt uncommitted to a particular path, we unlock the AI’s ability to surface new paths and weigh them against each other. |
If you want a copy-paste starting point, this prompt works in ChatGPT or Claude with no special tools: |
"I want [the outcome], by [when]. Here's my situation: [your constraints, your resources]. What are all my options? Lay out the trade-offs of each one in a neat table. Analysis only.” |
|
|
|
|
Strong Models Still Need Context |
If outcome-first prompting is this useful, why weren't we all doing it two years ago? |
Because two years ago, it didn't work very well. Asking an open-ended "what are my options" question returned generic output that was novel (at the time) but rarely useful without a couple rounds of clarification. |
That constraint has lifted. You don't have to change tools or learn a new technique to collect this upgrade: all the latest flagship models from Anthropic’s Claude and OpenAI’s ChatGPT are more than enough. The capability is already sitting in the chat window you use every day. |
As always, don’t forget to supply as much relevant context about your task as you can. |
A smart model with no idea what you're actually trying to achieve is still just guessing. When provided with enough context, the model searches for paths that fit your world, with your deadline and your constraints, instead of paths that fit a hypothetical average person. |
The next time you find yourself faced with a project you don’t have a ton of domain expertise in, consider using outcome-first prompting to make sure you’re pursuing the right path -- before you pour hours of effort into it. |
And if you’ve discovered any similar prompting unlocks from your own experimentation, I’d love to hear them. Hit reply on this email and tell me all about them. I read every email, and reply to a few where I can 🙂 |
—Dharmesh (@dharmesh) |
|
|
|
|
What'd you think of today's email?Click below to let me know. |
|
|
