The GRACE Framework: How to Write AI Prompts That Actually Work
>Your AI answers aren't mediocre because the tool is bad. It's almost always the prompt. Amber Boaz has a five-part framework to fix that, and she runs it live.
Most people think AI gives them mediocre answers because the tool isn't that good.
It's almost always the prompt.
That's the argument Amber Boaz makes on episode 2 of the Revenue Engineer Podcast, and she came with a framework to fix it. It's called GRACE: Goal, Role, Action, Context, Example. Five things to put in a prompt so the AI actually knows what you want.
Amber is a technical program manager and a Dreamforce speaker, and she built GRACE in collaboration with Claude. We didn't just talk about it. She ran it live, going from a blank cursor to a polished LinkedIn post, then to a full Salesforce UAT test plan, in real time.
Here's the framework, the live builds, and the handful of moves that take you from "the AI helped" to "the AI does real work."
GRACE, one letter at a time
The whole framework is five questions you answer before you hit enter:
- G is for Goal.What outcome do you want? A bulleted list, a first draft, a decision, a plan. Sometimes the goal is just to think out loud with a partner, and that's a valid goal too. If you don't name the destination, the AI picks one for you, and it usually picks "a wall of text."
- R is for Role.Who should the AI act as? A digital marketer, a Salesforce admin, a tough editor, a patient teacher. The same question gets a very different answer depending on who you tell it to be. Amber's favorite for hard topics: "explain this to me like I'm five."
- A is for Action.What exact task do you want done right now? Goal is the outcome, action is the move that gets you there. "Help me with my email" is vague. "Write a first draft of this email" is something the AI can actually finish.
- C is for Context.What background does it need? This is the piece almost everyone skips, and it's where all the leverage is. Why is this email hard. What are the field names. Who is the audience. Paste the job description, paste the error message, paste the profile.
- E is for Example. What does good look like? Show it a past post you liked or a format you want to copy, and it matches the pattern.
Watching it work: two live builds
Theory is easy. The reason this episode is worth your time is that Amber put GRACE on screen and let it run.
Build one: a LinkedIn post on Duck AI
She typed a real prompt using the framework. Role: an expert digital marketer who specializes in professional networking. Action: three short post options about an upcoming webinar. Then she iterated, the way you actually should. No em dashes. Less salesy. Soften that hook. The first answer was the start of the conversation, not the end of it.
Build two: a Salesforce UAT test plan, from a few sentences
This is the one that lands for the practitioners. We described a small app live (a ZIP code field that calls a weather API and writes the result back to the record), told the AI to ask us clarifying questions, and answered them.
What came back was a real test plan: test types, scope and priorities, the happy path, data integrity checks, invalid ZIP codes, cross-browser cases, environment data, even a one-week test schedule. The kind of document that, on a consulting project, you either build from scratch under pressure or beg off a busy senior who hands you a five-year-old example and says "figure it out."
Five moves from the episode
The framework gets you started. These are the habits that make the difference.
- Iterate. Never one-shot.GRACE opens the conversation, it doesn't close it. Read the first answer, tell it what to fix, run it again. The magic is in the back-and-forth.
- Teach it your voice.Paste in a few of your old posts, emails, or notes and say "this is how I write." Amber fed Claude her past blog posts and her exported tweets. This is the fix for "everything AI writes feels fake."
- Add about 5% human mess on purpose. Perfect grammar reads as a robot. Ask for a few small, natural imperfections and the writing sounds like a person typed it.
- Let the AI write your prompt.Stuck on how to ask? Tell it: "become my prompt engineer, ask me questions, then write the best version of this prompt." You get a strong, reusable prompt you can save. People call this meta prompting. It works.
- Keep a human in the loop. The AI is fast, not flawless. It makes assumptions and sometimes gets things wrong. Read the output before it goes to a customer, a VP, or production.
Try it in the next 60 seconds
Here's the template. Copy it, fill in the blanks, paste it into Claude, ChatGPT, Duck AI, whatever you have open:
Goal: I want to end up with ________________________. Role: Act as ________________________. Action: ________________________ (use a clear verb: write, list, compare, draft, rewrite, build, summarize). Context: Here is the background you need: ________________________. The audience is ________________________. The tone is ______________. Example: Here is what good looks like: ________________________.
Pick one real task you've been putting off. Fill in Goal, Action, and Context at minimum. Send it. Then reply with one correction and watch it get better. That's the whole habit.
Watch the full episode
Revenue Engineer Episode 2 with Amber Boaz is on the walters954 YouTube channel. This is part of an ongoing series where real practitioners show their AI-assisted workflow on camera, with the rough edges left in.
Building at the intersection of Salesforce and AI?
If this format is useful to you, subscribe on YouTube and sign up for the newsletter at revenueeng.com. New episodes and written breakdowns go there first.
Revenue Engineer is the community for working Salesforce professionals leveling up on AI-augmented development and modern GTM tooling. Weekly drops, live sessions, and a founding cohort opening soon. Join the waitlist and you'll be first in line.
Follow Amber on LinkedIn. If you try GRACE on something real, drop a comment on the YouTube video and tell me what you built. I read every one.
// one email, no spam, first in line.